From ed8f7ae977cb392b24b9038a55c5c82454cb4e5f Mon Sep 17 00:00:00 2001
From: billowliu2 <billowliu2@users.noreply.github.com>
Date: Sun, 2 Aug 2026 01:24:28 +0800
Subject: [PATCH] =?UTF-8?q?=E5=88=9D=E5=A7=8B=E5=8C=96=20kimi-eyes?=
 =?UTF-8?q?=EF=BC=9A=E7=BB=99=20KimiCode=20=E9=9D=9E=E5=A4=9A=E6=A8=A1?=
 =?UTF-8?q?=E6=80=81=E6=A8=A1=E5=9E=8B=E8=A1=A5=E4=B8=8A=E8=A7=86=E8=A7=89?=
 =?UTF-8?q?=E8=83=BD=E5=8A=9B?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .gitignore              |  13 ++
 LICENSE                 |  21 +++
 README.en.md            | 268 +++++++++++++++++++++++++++++++
 README.md               | 198 +++++++++++++++++++++++
 SYSTEM.md               |  27 ++++
 kimi.plugin.json        |  21 +++
 mcp/models-db.json      |   1 +
 mcp/server.mjs          | 241 ++++++++++++++++++++++++++++
 mcp/vision.mjs          | 344 ++++++++++++++++++++++++++++++++++++++++
 package.json            |  40 +++++
 scripts/sync-models.mjs |  95 +++++++++++
 setup.mjs               | 241 ++++++++++++++++++++++++++++
 12 files changed, 1510 insertions(+)
 create mode 100644 .gitignore
 create mode 100644 LICENSE
 create mode 100644 README.en.md
 create mode 100644 README.md
 create mode 100644 SYSTEM.md
 create mode 100644 kimi.plugin.json
 create mode 100644 mcp/models-db.json
 create mode 100644 mcp/server.mjs
 create mode 100644 mcp/vision.mjs
 create mode 100644 package.json
 create mode 100644 scripts/sync-models.mjs
 create mode 100644 setup.mjs

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..5d8b796
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,13 @@
+# dependencies
+node_modules/
+
+# npm publish artifacts
+*.tgz
+
+# local secrets — never commit (credentials live in user-level ~/.npmrc)
+.env
+.env.*
+
+# local verification / scratch
+.verify/
+*.log
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000..1cf8e2b
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 billowliu2
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.en.md b/README.en.md
new file mode 100644
index 0000000..2c5f76c
--- /dev/null
+++ b/README.en.md
@@ -0,0 +1,268 @@
+# Kimi Eyes 👀
+
+> [中文](README.md) | English
+
+Give **non-multimodal models** in Kimi Code the ability to analyze images and
+screenshots (inspired by [opencode-vision](https://github.com/JochenYang/opencode-vision),
+but thinner: no hooks, no message transforms, no leftover state).
+
+**How it works**: the plugin declares an MCP stdio server exposing two tools —
+`read_image` (read a local image) and `read_clipboard_image` (read a clipboard
+screenshot) — and a `SYSTEM.md` guide tells the model when to call them. The server
+sends the image to your own vision API (OpenAI-compatible or Anthropic protocol) and
+returns the text description.
+
+```
+Multimodal models: paste with Alt+V, see natively — the plugin stays idle.
+Non-multimodal models: @image-path / "analyze this screenshot"
+  → model calls mcp__kimi-eyes__read_image / read_clipboard_image
+  → your VLM returns a description
+```
+
+## Prerequisites
+
+- Kimi Code CLI (with `/plugins` and MCP support)
+- Node.js ≥ 18 (`node --version`; native `fetch` requires 18+)
+- A multimodal vision API of your own (OpenAI-compatible `chat/completions`, or
+  Anthropic `messages`) — you provide the key
+
+## Compatibility
+
+**The plugin puts no restriction on your main model.** No matter which model your
+Kimi Code session is running, or which provider it comes from, as long as it lacks
+native image input (`image_in`), this plugin adds vision to it:
+
+- **Common non-multimodal models**: DeepSeek (`deepseek-chat`), Qwen text-only
+  (`qwen-plus` / `qwen-turbo`), Llama text variants, and locally deployed text
+  models (Ollama / vLLM, etc.)
+- **Any third-party model**: any OpenAI-compatible or custom model configured in
+  Kimi Code — if it cannot see images natively, the plugin guides it to call the
+  vision tools whenever an image is involved
+- **Multimodal models**: skipped automatically (paste with Alt+V, see natively)
+
+Vision capability is provided by the **VLM you configure**, fully decoupled from
+the main model:
+
+| Protocol | Example vision models |
+| --- | --- |
+| OpenAI-compatible | qwen-vl series, GLM-4V, GPT-4o / GPT-5 compatible endpoints, Gemini-compatible endpoints, etc. |
+| Anthropic | Claude 3.5 / 3.7 / 4 series, etc. |
+
+In short: **the main model handles text, the external VLM handles images.** When
+configuring, just pick the protocol that matches your vision API provider (step 1
+of the setup wizard); everything else is automatic.
+
+## Quick start
+
+### 1. Install the plugin
+
+In a Kimi Code session:
+
+```
+/plugins install D:\AIGC\Plugin\kimi-eyes
+```
+
+### 2. Configure the vision API (one-time)
+
+Run the setup wizard from the plugin directory:
+
+```
+cd D:\AIGC\Plugin\kimi-eyes
+node setup.mjs
+```
+
+The wizard walks you through: **choose protocol (OpenAI-compatible / Anthropic) →
+enter Base URL → enter API Key (masked input) → fetch model list and pick a
+multimodal model → 1×1 image vision check → optionally enter your main model name
+(for trigger decisions) → write config**.
+
+- The model list is fetched from `GET {BaseUrl}/models`; entries are tagged
+  **"✓vision / text"** using the bundled models.dev database, with a "★guess" fallback
+  for models the database does not know; if fetching fails it falls back to manual entry
+- The vision check must pass before the config is saved, so you never end up with a
+  model that rejects images
+- The last step asks for your **main model name** (optional): the plugin looks it up in
+  models.dev, and if that model supports native image input the tools refuse with a
+  hint — the code-level "multimodal → don't trigger" switch (see
+  [Triggering](#triggering-everyday-fully-automatic))
+- Config is written to `~/.kimi-code/kimi-eyes/config.json` (`chmod 600` on
+  non-Windows systems)
+
+### 3. Enable
+
+```
+/reload
+```
+
+The MCP server starts automatically with the session.
+
+### 4. Use it
+
+| Scenario | Action |
+| --- | --- |
+| Multimodal model | Paste with Alt+V directly — native vision, plugin not involved |
+| Non-multimodal + image path | Type `@screenshot.png` or paste the path; the model calls `read_image` |
+| Non-multimodal + just screenshotted/copied | Ask "analyze this screenshot"; the model calls `read_clipboard_image` to read the system clipboard |
+
+## Activation and triggering
+
+### Activation (one-time)
+
+```
+/plugins install D:\AIGC\Plugin\kimi-eyes   # 1. Install
+/reload                                     # 2. Enable (or start a new session with /new)
+```
+
+Once enabled, the MCP server starts automatically with every session — there is no
+separate "turn on the feature" step. To verify:
+
+- `/plugins list` → kimi-eyes should show as enabled
+- `/mcp` → the kimi-eyes server should show as connected
+- `/plugins info kimi-eyes` → should show no diagnostics errors
+
+### Triggering (everyday, fully automatic)
+
+The plugin is **passive**: `SYSTEM.md` plants the rules into the model, and the model
+calls the tools automatically at the right moment — you do nothing:
+
+| Signal (anything in the message) | Model's automatic behavior |
+| --- | --- |
+| Image-format path or `@` reference (`.png/.jpg/.jpeg/.webp/.gif/.bmp`) | **Hard trigger**: unconditionally calls `read_image(path)`, regardless of wording |
+| Media content you cannot interpret (e.g. a pasted image) | Ignores that media part, calls `read_clipboard_image()` (the pasted image is almost always still in the clipboard) |
+| Wording implies image content: image / screenshot / photo / UI / chart / CAPTCHA / OCR, etc., but no path | Calls `read_clipboard_image()` |
+| Model natively supports `image_in` (multimodal) | Skips all rules, sees images natively, never calls the tools |
+
+**Triggering does not depend on fixed wording** — the user does not need to say
+"analyze the image". Image-format paths are an **unconditional hard trigger**; when
+in doubt, the model is instructed to call a tool rather than guess.
+
+**On model-capability detection**: Kimi Code does not expose "is the current model
+multimodal?" to plugins, so this plugin provides two layers:
+
+- **Code-level (recommended)**: declare `mainModel` in the config (last wizard step,
+  or the `VISION_MAIN_MODEL` environment variable). The plugin looks it up in the
+  bundled models.dev database — if it supports image input, the tools refuse with a
+  hint ("paste the image directly"); if it is text-only, the tools proceed normally.
+- **Prompt-level**: without `mainModel`, the SYSTEM.md explicit skip rule handles it
+  (the model recognizes its own capability).
+
+Both layers are fail-safe: a wrong judgment costs at most one wasted external call
+(a multimodal model calling a tool) or a missed trigger (a text model — covered by
+the path/semantic signals). For a hard switch, simply `/plugins disable kimi-eyes`
+when running multimodal models.
+
+**The first tool call prompts one approval** (MCP tool permission): choose *Approve
+for this session* to skip prompts for the rest of the session; for permanent
+approval, add to `~/.kimi-code/config.toml`:
+
+```toml
+[[permission.rules]]
+decision = "allow"
+pattern = "mcp__kimi-eyes__*"
+```
+
+A successful trigger looks like: a tool call appears in the TUI before the answer,
+and the model then answers based on the returned description. If the model does not
+call the tool on its own (e.g. you pasted a path without asking a question), just
+command it: "Call the read_image tool to analyze D:\xxx.png".
+
+## Environment variables (optional; override the config file)
+
+| Variable | Description | Priority |
+| --- | --- | --- |
+| `VISION_API_PROTOCOL` | `openai` or `anthropic`; force the protocol | Higher than config file |
+| `VISION_API_KEY` | API key | Higher than config file |
+| `VISION_API_URL` | Base URL | Higher than config file |
+| `VISION_MODEL` | Model name | Higher than config file |
+| `VISION_MAIN_MODEL` | Your main model name (optional), used for trigger decisions | Higher than config file |
+| `VISION_MAX_TOKENS` | Max tokens for the vision response (default 1024) | — |
+| `VISION_FETCH_TIMEOUT_MS` | Request timeout in ms (default 60000) | — |
+
+> When an environment variable conflicts with `config.json`, the variable wins. If
+> neither is set, the tools return a clear error pointing at `setup.mjs`.
+> Variable names are compatible with opencode-vision, so migrating is trivial.
+
+## Protocol details
+
+| | OpenAI-compatible | Anthropic |
+| --- | --- | --- |
+| Request endpoint | `{BaseUrl}/chat/completions` (`/v1` auto-appended) | `{BaseUrl}/v1/messages` |
+| Auth | `Authorization: Bearer <Key>` | `x-api-key: <Key>` + `anthropic-version: 2023-06-01` |
+| Image payload | `image_url` + base64 data URL | `source: {type:"base64"}` |
+| Model list | `GET {BaseUrl}/models` | `GET {BaseUrl}/v1/models` (not an official Anthropic endpoint; falls back to manual entry) |
+
+## Model capability database (models.dev)
+
+The plugin ships a slim capability cache `mcp/models-db.json` synced from
+[models.dev](https://models.dev) (currently 279+ models, tagged with whether each
+supports image input, ~19 KB). It powers two things:
+
+- **Exact tagging** in the setup wizard's model picker ("✓vision / text"), replacing
+  pure keyword guessing
+- The **code-level trigger switch**: once `mainModel` is declared, the plugin checks
+  whether your main model is multimodal
+
+**Sync at packaging time** (the data evolves — run before each release):
+
+```
+node scripts/sync-models.mjs
+```
+
+- Source: `https://models.dev/models.json`
+- Behind a proxy: set `HTTPS_PROXY`, e.g. `HTTPS_PROXY=http://127.0.0.1:7897 node scripts/sync-models.mjs`
+- Options: `--timeout <seconds>` (default 180), `--out <path>` (default
+  `mcp/models-db.json`), `--endpoint <URL>`
+- Matching strategy: exact id → `provider/model-name` suffix → case-insensitive name
+- **Unknown models**: lookup returns unknown — the wizard falls back to keyword
+  guessing ("★guess") and triggering falls back to the SYSTEM.md rules; nothing breaks
+- **Manual extras**: models.dev does not list every vision model (e.g. `k3-256k`,
+  `kimi-for-coding`). Keep them in `EXTRA_ENTRIES` inside
+  `scripts/sync-models.mjs` — every sync merges them in, so re-syncs never drop them
+
+## Tools
+
+| Tool | Arguments | Description |
+| --- | --- | --- |
+| `read_image` | `path` (required), `prompt` (optional) | Validates the file is an image (extension + magic bytes), then calls the VLM |
+| `read_clipboard_image` | `prompt` (optional) | Captures the clipboard image to a temp file, calls the VLM, deletes the temp file |
+
+Clipboard capture depends on the platform: Windows uses PowerShell (built-in),
+macOS needs `pngpaste` (`brew install pngpaste`), Linux needs `wl-paste` (Wayland)
+or `xclip` (X11).
+
+## Troubleshooting
+
+- **Tool returns "Vision API is not configured"** → run `node setup.mjs`, or set
+  `VISION_API_KEY` / `VISION_API_URL` / `VISION_MODEL`
+- **Model list fetch fails** (404/401) → the wizard falls back to manual entry; if
+  your provider has no `/models` endpoint, just type the model name
+- **Vision check fails** → pick a model that really accepts image input (e.g.
+  `qwen-vl-max`, `glm-4v`, `gpt-4o`, `claude-3-5-sonnet` — check your provider's docs)
+- **`read_clipboard_image` errors** → make sure the clipboard actually holds an image
+  (Ctrl+C an image or Win+Shift+S a screenshot first); on macOS/Linux check the
+  platform tool above is installed
+- **Tools not showing up after install** → verify the plugin is enabled
+  (`/plugins list`) and run `/reload` or start a new session
+
+## Security notes
+
+- `config.json` stores your API key in plain text (`600` perms on non-Windows);
+  never commit it to a repository
+- `read_clipboard_image` reads the system clipboard — it may contain sensitive
+  content you just copied. The tool call goes through the approval flow, so you
+  decide when it runs
+- This project ships no credentials; vision requests go only to the Base URL you
+  configured
+
+## Limitations
+
+- No subagent delegation: Kimi Code's `model_preference` only supports
+  primary/secondary (it cannot name a specific vision model the way opencode can),
+  so the value is limited
+- No `UserPromptSubmit` hook fallback: the `SYSTEM.md` guide covers the common
+  cases; if paste/reference behavior misbehaves in your TUI, a hook can be
+  re-evaluated then
+
+## License
+
+MIT
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..1e8f622
--- /dev/null
+++ b/README.md
@@ -0,0 +1,198 @@
+# Kimi Eyes 👀
+
+> [English](README.en.md) | 中文
+
+让 KimiCode 的**非多模态模型**也能分析图片与截图（借鉴 [opencode-vision](https://github.com/JochenYang/opencode-vision) 的思路，但更薄：无 hook、无消息变换、无状态残留）。
+
+**核心机制**：插件声明一个 MCP stdio 服务器，暴露两个工具——`read_image`（读本地图片）与 `read_clipboard_image`（读剪贴板截图），经 `SYSTEM.md` 引导模型调用；服务器按你配置的协议（OpenAI 兼容 / Anthropic）把图片发给你自己配置的多模态 API，返回文本描述。
+
+```
+多模态模型：Alt+V 粘贴即看，插件自动闲置
+非多模态模型：@图片路径 / 截图后提问 → 模型调 mcp__kimi-eyes__read_image / read_clipboard_image → 你的 VLM 返回描述
+```
+
+## 前置要求
+
+- KimiCode CLI（本插件通过插件机制加载，需支持 `/plugins` 与 MCP）
+- Node.js ≥ 18（`node --version` 检查；原生 fetch 需要 18+）
+- 一个支持视觉输入的多模态 API（OpenAI 兼容 `chat/completions`，或 Anthropic `messages`），Key 由你自己提供
+
+## 适用范围
+
+**本插件对主模型没有任何限制**——无论你的 KimiCode 当前用哪个模型、来自哪家服务商，只要它不具备原生图片输入能力（`image_in`），本插件就能为它补上视觉：
+
+- **常见非多模态模型**：DeepSeek（deepseek-chat）、Qwen 纯文本版（qwen-plus / qwen-turbo）、Llama 文本系列、以及本地部署的文本模型（Ollama / vLLM 等）
+- **任何第三方模型**：通过 KimiCode 配置的任意 OpenAI 兼容或自定义模型，只要它原生不看图，插件就会在消息含图片时引导模型调用视觉工具
+- **多模态模型**：自动跳过（Alt+V 粘贴即看，插件闲置）
+
+视觉能力由**你自己配置的 VLM** 提供，与主模型完全解耦：
+
+| 协议 | 常见视觉模型示例 |
+| --- | --- |
+| OpenAI 兼容 | qwen-vl 系列、GLM-4V、GPT-4o / GPT-5 兼容端点、Gemini 兼容端点等 |
+| Anthropic | Claude 3.5 / 3.7 / 4 系列等 |
+
+一句话：**主模型只管理解文字，看图交给外部 VLM**。配置时只需在 setup 向导第一步选择与你视觉 API 服务商匹配的协议，其余全部自动。
+
+## 快速开始
+
+### 1. 安装插件
+
+在 KimiCode 会话里执行：
+
+```
+/plugins install D:\AIGC\Plugin\kimi-eyes
+```
+
+### 2. 配置视觉 API（一次性）
+
+在插件目录运行配置向导：
+
+```
+cd D:\AIGC\Plugin\kimi-eyes
+node setup.mjs
+```
+
+向导按顺序引导：**选择协议（OpenAI 兼容 / Anthropic）→ 填写 BaseUrl → 填写 API Key（掩码输入）→ 拉取模型列表选择多模态模型 → 1×1 图片视觉验证 → 可选填写当前主模型名（用于触发判断）→ 写入配置**。
+
+- 模型列表自动从 `GET {BaseUrl}/models` 拉取，用内置的 models.dev 数据库**精确标注「✓视觉 / 文本」**，未收录的按关键词标「★疑似」；拉取失败时回退为手动输入
+- 视觉验证通过才落盘，防止配到一个不收图片的模型
+- 最后一步可填写**当前主模型名**（可选）：插件用 models.dev 判定后，若该模型支持原生识图，工具会直接提示无需调用——即「多模态不触发」的代码级判断（详见[触发](#触发日常全自动)）
+- 配置写入 `~/.kimi-code/kimi-eyes/config.json`（类 Unix 系统自动 `chmod 600`）
+
+### 3. 启用
+
+```
+/reload
+```
+
+MCP 服务器会随会话自动启动。
+
+### 4. 使用
+
+| 场景 | 操作 |
+| --- | --- |
+| 多模态模型 | 直接 Alt+V 粘贴图片，原生看图，插件不参与 |
+| 非多模态 + 有图片路径 | 输入 `@截图.png` 或直接给路径，模型自动调 `read_image` |
+| 非多模态 + 刚截图/复制 | 截图后直接提问「分析这张截图」，模型自动调 `read_clipboard_image` 读系统剪贴板 |
+
+## 激活与触发
+
+### 激活（一次性）
+
+```
+/plugins install D:\AIGC\Plugin\kimi-eyes   # 1. 安装
+/reload                                     # 2. 启用（或 /new 新开会话）
+```
+
+启用后 MCP 服务器随每个会话自动启动，没有单独的「开启功能」步骤。验证方式：
+
+- `/plugins list` → kimi-eyes 状态应为 enabled
+- `/mcp` → kimi-eyes 服务器应显示 connected
+- `/plugins info kimi-eyes` → 应无诊断错误
+
+### 触发（日常，全自动）
+
+插件是**被动式**的：`SYSTEM.md` 为模型植入规则，模型在合适时机自动调用工具，无需手动操作：
+
+| 信号（消息里的任何一项） | 模型自动行为 |
+| --- | --- |
+| 图片格式路径或 `@` 引用（`.png/.jpg/.jpeg/.webp/.gif/.bmp`） | **硬触发**：无条件调 `read_image(path)`，与提问语料无关 |
+| 消息里有你无法解读的媒体内容（如粘贴的图片） | 忽略该媒体 part，自动调 `read_clipboard_image()` 读剪贴板（粘贴后剪贴板仍保留原图） |
+| 提问语义涉及图像内容：图 / 截图 / 照片 / 界面 / 图表 / 验证码 / OCR 等，但无路径 | 自动调 `read_clipboard_image()` |
+| 模型原生支持 `image_in`（多模态） | 跳过全部规则，原生看图，不调工具 |
+
+**触发不依赖固定语料**——用户不需要说「分析这个图片」。图片格式路径是**无条件硬触发**；拿不准时模型会倾向于调工具而不是瞎猜（SYSTEM.md 里的明确规则）。
+
+**关于模型能力判断**：KimiCode 不向插件暴露「当前模型是否多模态」的信号，本插件提供两层判断：
+
+- **代码级（推荐）**：在配置里声明 `mainModel`（setup 向导最后一步，或环境变量 `VISION_MAIN_MODEL`），插件用内置的 models.dev 数据库判断——命中且支持识图 → 工具直接提示「你是多模态模型，直接粘贴看图」；命中且纯文本 → 正常走工具
+- **引导级**：未声明 `mainModel` 时，靠 SYSTEM.md 的显式跳过规则（模型自我识别）
+
+两层都是 fail-safe：判断错误最坏只是多一次外部调用（多模态误调）或漏调（文本模型漏调可用路径/语义信号补上）。若想硬性关闭，多模态场景可直接 `/plugins disable kimi-eyes`。
+
+**首次调用会弹一次审批**（MCP 工具权限）：选 *Approve for this session* 本会话免问；想永久免审批，在 `~/.kimi-code/config.toml` 添加：
+
+```toml
+[[permission.rules]]
+decision = "allow"
+pattern = "mcp__kimi-eyes__*"
+```
+
+触发成功的标志：回答前 TUI 出现工具调用记录，模型随后基于返回的描述作答。若模型未自动调用（例如只贴了路径未提问），可直接命令：「调用 read_image 工具分析 D:\xxx.png」。
+
+## 环境变量（可选，覆盖配置文件）
+
+| 变量 | 说明 | 优先级 |
+| --- | --- | --- |
+| `VISION_API_PROTOCOL` | `openai` 或 `anthropic`，强制指定协议 | 高于配置文件 |
+| `VISION_API_KEY` | API Key | 高于配置文件 |
+| `VISION_API_URL` | BaseUrl | 高于配置文件 |
+| `VISION_MODEL` | 模型名 | 高于配置文件 |
+| `VISION_MAIN_MODEL` | 当前主模型名（可选），用于触发判断 | 高于配置文件 |
+| `VISION_MAX_TOKENS` | 视觉 API 返回最大 token 数（默认 1024） | — |
+| `VISION_FETCH_TIMEOUT_MS` | 请求超时毫秒（默认 60000） | — |
+
+> 环境变量与 `config.json` 同名配置冲突时，环境变量优先；两者都没配时工具会返回明确错误并提示先运行 `setup.mjs`。
+> 环境变量可沿用 opencode-vision 的迁移习惯——变量名完全兼容。
+
+## 协议说明
+
+| | OpenAI 兼容 | Anthropic |
+| --- | --- | --- |
+| 请求端点 | `{BaseUrl}/chat/completions`（自动补 `/v1`） | `{BaseUrl}/v1/messages` |
+| 鉴权 | `Authorization: Bearer <Key>` | `x-api-key: <Key>` + `anthropic-version: 2023-06-01` |
+| 图片传法 | `image_url` + base64 data URL | `source: {type:"base64"}` |
+| 模型列表 | `GET {BaseUrl}/models` | `GET {BaseUrl}/v1/models`（官方无此端点，失败回退手动输入） |
+
+## 模型能力数据库（models.dev）
+
+插件内置一份从 [models.dev](https://models.dev) 同步的精简能力库 `mcp/models-db.json`（当前收录 279+ 模型，标注每个模型是否支持图片输入，约 19KB），用于两处：
+
+- 向导选择视觉模型时的**精确标注**（✓视觉 / 文本），替代纯关键词猜测
+- 声明 `mainModel` 后**判断当前主模型是否多模态**，实现代码级触发开关
+
+**打包时同步**（数据会更新，建议每次发布前运行）：
+
+```
+node scripts/sync-models.mjs
+```
+
+- 数据源：`https://models.dev/models.json`
+- 走代理：设置 `HTTPS_PROXY`，例如 `HTTPS_PROXY=http://127.0.0.1:7897 node scripts/sync-models.mjs`
+- 可选参数：`--timeout <秒>`（默认 180）、`--out <路径>`（默认 `mcp/models-db.json`）、`--endpoint <URL>`
+- 模型匹配策略：精确 id → `provider/模型名` 后缀 → 大小写不敏感的名称匹配
+- **未收录的模型**：查询返回未知——向导回退关键词猜测（★疑似），触发判断回退 SYSTEM.md 引导，不影响使用
+- **手动特例**：models.dev 未收录但确认支持识图的模型（如 `k3-256k`、`kimi-for-coding`），在 `scripts/sync-models.mjs` 的 `EXTRA_ENTRIES` 里维护，每次同步自动合入、不会被覆盖
+
+## 工具
+
+| 工具 | 参数 | 说明 |
+| --- | --- | --- |
+| `read_image` | `path`（必填）、`prompt`（可选） | 校验文件为图片（扩展名 + 魔数）后调 VLM |
+| `read_clipboard_image` | `prompt`（可选） | 从系统剪贴板取图存临时文件后调 VLM，用完即删 |
+
+剪贴板取图依赖平台工具：Windows 用 PowerShell（内置）、macOS 需 `pngpaste`（`brew install pngpaste`）、Linux 需 `wl-paste`（Wayland）或 `xclip`（X11）。
+
+## 故障排查
+
+- **工具返回「Vision API is not configured」** → 运行 `node setup.mjs`，或设置 `VISION_API_KEY` / `VISION_API_URL` / `VISION_MODEL`
+- **模型列表拉取失败**（404/401）→ 向导自动回退手动输入；若服务端不支持 `/models`，直接输入模型名即可
+- **视觉验证失败** → 换一个真正支持图片输入的模型（参考服务商文档，如 `qwen-vl-max`、`glm-4v`、`gpt-4o`、`claude-3-5-sonnet` 等）
+- **`read_clipboard_image` 报错** → 确认剪贴板里确实有图片（先 Ctrl+C 复制或 Win+Shift+S 截图）；macOS/Linux 检查上述平台工具是否安装
+- **安装后工具不生效** → 确认插件已启用（`/plugins list`），并运行过 `/reload` 或新开会话
+
+## 安全说明
+
+- `config.json` 以明文保存你的 API Key（类 Unix 已设 `600` 权限）；请勿将该文件提交到版本库
+- `read_clipboard_image` 会读取系统剪贴板内容——剪贴板里可能刚复制过敏感信息。该工具调用会经过审批流，由你决定何时触发
+- 本项目不内置任何密钥，视觉请求只发往你配置的 BaseUrl
+
+## 局限
+
+- 不做子代理委托方案：KimiCode 子代理的 `model_preference` 只能选择 primary/secondary，无法像 opencode 那样指定一个具体视觉模型，收益有限
+- 不做 `UserPromptSubmit` hook 兜底：当前 `SYSTEM.md` 引导已覆盖常规场景；若你的 TUI 环境下粘贴/引用行为异常，可再评估 hook
+
+## License
+
+MIT
diff --git a/SYSTEM.md b/SYSTEM.md
new file mode 100644
index 0000000..a39e9ac
--- /dev/null
+++ b/SYSTEM.md
@@ -0,0 +1,27 @@
+# Kimi Eyes — Vision Assist Rules
+
+**If your model supports image input (`image_in`) natively, ignore every rule below —
+you can see images directly, so never call these tools.**
+
+For models **without** native image input, the signals below are **hard triggers**:
+call a vision tool whenever any of them appears, regardless of how the user phrased
+their request.
+
+1. **Image-format path / `@` reference (hard trigger, unconditional)** — any path
+   ending in `.png`, `.jpg`, `.jpeg`, `.webp`, `.gif`, or `.bmp` (including an `@`
+   file reference to such a file) → call `mcp__kimi-eyes__read_image` with that
+   path. Pass `prompt` when the user has a specific question.
+2. **Media content you cannot interpret** — a pasted image or any media part you
+   cannot read → ignore that part and call
+   `mcp__kimi-eyes__read_clipboard_image` (the image is almost always still in the
+   system clipboard).
+3. **Wording implies an image, no path** — the user refers to "this", "the
+   screenshot", "the UI/interface", "the chart", "the photo", asks for OCR / CAPTCHA
+   reading, or otherwise implies image content, with no path and no visible media
+   part → call `mcp__kimi-eyes__read_clipboard_image`.
+
+When in doubt, call a tool rather than guessing blindly about the image.
+
+If a tool reports that the vision API is not configured, tell the user to run
+`node setup.mjs` inside the kimi-eyes plugin directory (or set the
+`VISION_API_KEY` / `VISION_API_URL` / `VISION_MODEL` environment variables).
diff --git a/kimi.plugin.json b/kimi.plugin.json
new file mode 100644
index 0000000..5e006e3
--- /dev/null
+++ b/kimi.plugin.json
@@ -0,0 +1,21 @@
+{
+  "name": "kimi-eyes",
+  "version": "1.0.0",
+  "description": "Give non-multimodal models vision: analyze local images and clipboard screenshots through a user-configured VLM API (OpenAI-compatible or Anthropic protocol).",
+  "keywords": ["vision", "image", "screenshot", "multimodal", "vlm"],
+  "author": "kimi-eyes",
+  "license": "MIT",
+  "interface": {
+    "displayName": "Kimi Eyes",
+    "shortDescription": "视觉辅助：让非多模态模型也能分析图片/截图",
+    "longDescription": "读取本地图片或剪贴板截图，调用用户自配的多模态 API（OpenAI 兼容或 Anthropic 协议）返回描述。多模态模型不受影响，粘贴即原生看图。"
+  },
+  "systemPromptPath": "./SYSTEM.md",
+  "mcpServers": {
+    "kimi-eyes": {
+      "command": "node",
+      "args": ["./mcp/server.mjs"],
+      "cwd": "./"
+    }
+  }
+}
diff --git a/mcp/models-db.json b/mcp/models-db.json
new file mode 100644
index 0000000..a5adbe2
--- /dev/null
+++ b/mcp/models-db.json
@@ -0,0 +1 @@
+{"zhipuai/glm-4.6v":{"name":"GLM-4.6V","image":true},"zhipuai/glm-5":{"name":"GLM-5","image":false},"zhipuai/glm-4.5-air":{"name":"GLM-4.5-Air","image":false},"zhipuai/glm-5.1":{"name":"GLM-5.1","image":false},"zhipuai/glm-4.7-flash":{"name":"GLM-4.7-Flash","image":false},"zhipuai/glm-5.2":{"name":"GLM-5.2","image":false},"zhipuai/glm-4.7-flashx":{"name":"GLM-4.7-FlashX","image":false},"zhipuai/glm-4.6":{"name":"GLM-4.6","image":false},"zhipuai/glm-4.5":{"name":"GLM-4.5","image":false},"zhipuai/glm-4.5v":{"name":"GLM-4.5V","image":true},"zhipuai/glm-4.7":{"name":"GLM-4.7","image":false},"zhipuai/glm-5-turbo":{"name":"GLM-5-Turbo","image":false},"zhipuai/glm-5v-turbo":{"name":"GLM-5V-Turbo","image":true},"zhipuai/glm-4.5-flash":{"name":"GLM-4.5-Flash","image":false},"microsoft/mai-code-1-flash":{"name":"MAI-Code-1-Flash","image":false},"cohere/command-r7b-arabic-02-2025":{"name":"Command R7B Arabic","image":false},"cohere/command-r-08-2024":{"name":"Command R","image":false},"cohere/command-a-plus-05-2026":{"name":"Command A Plus","image":true},"cohere/command-a-translate-08-2025":{"name":"Command A Translate","image":false},"cohere/c4ai-aya-expanse-8b":{"name":"Aya Expanse 8B","image":false},"cohere/c4ai-aya-vision-32b":{"name":"Aya Vision 32B","image":true},"cohere/command-a-03-2025":{"name":"Command A","image":false},"cohere/command-r-plus-08-2024":{"name":"Command R+","image":false},"cohere/command-a-reasoning-08-2025":{"name":"Command A Reasoning","image":false},"cohere/command-a-vision-07-2025":{"name":"Command A Vision","image":true},"cohere/command-r7b-12-2024":{"name":"Command R7B","image":false},"cohere/c4ai-aya-vision-8b":{"name":"Aya Vision 8B","image":true},"cohere/north-mini-code-1-0":{"name":"North Mini Code","image":false},"cohere/c4ai-aya-expanse-32b":{"name":"Aya Expanse 32B","image":false},"nvidia/mistral-nemotron":{"name":"Mistral Nemotron","image":false},"nvidia/nemotron-nano-12b-v2-vl":{"name":"Nemotron Nano 12B v2 VL","image":true},"nvidia/nemotron-3-nano-30b-a3b":{"name":"Nemotron 3 Nano 30B A3B","image":false},"nvidia/llama-3.3-nemotron-super-49b-v1.5":{"name":"Llama 3.3 Nemotron Super 49B v1.5","image":false},"nvidia/llama-3.1-nemotron-safety-guard-8b-v3":{"name":"Llama 3.1 Nemotron Safety Guard 8B v3","image":false},"nvidia/llama-3.3-nemotron-super-49b-v1":{"name":"Llama 3.3 Nemotron Super 49B v1","image":false},"nvidia/llama-3.1-nemotron-70b-instruct":{"name":"Llama 3.1 Nemotron 70B Instruct","image":false},"nvidia/nemotron-3-super-120b-a12b":{"name":"Nemotron 3 Super 120B A12B","image":false},"nvidia/nemotron-3-content-safety":{"name":"Nemotron 3 Content Safety","image":false},"nvidia/nemotron-3-nano-omni-30b-a3b-reasoning":{"name":"Nemotron 3 Nano Omni 30B A3B Reasoning","image":true},"nvidia/llama-3.1-nemotron-ultra-253b":{"name":"Llama 3.1 Nemotron Ultra 253B","image":false},"nvidia/nemotron-voicechat":{"name":"Nemotron VoiceChat","image":false},"nvidia/llama-nemotron-embed-vl-1b-v2":{"name":"Llama Nemotron Embed VL 1B v2","image":true},"nvidia/nemotron-content-safety-reasoning-4b":{"name":"Nemotron Content Safety Reasoning 4B","image":false},"nvidia/nemotron-3-ultra-550b-a55b":{"name":"Nemotron 3 Ultra 550B A55B","image":false},"nvidia/llama-nemotron-rerank-vl-1b-v2":{"name":"Llama Nemotron Rerank VL 1B v2","image":true},"nvidia/nemotron-mini-4b-instruct":{"name":"Nemotron Mini 4B Instruct","image":false},"nvidia/nemotron-nano-9b-v2":{"name":"Nemotron Nano 9B v2","image":false},"nvidia/nemotron-cascade-2-30b-a3b":{"name":"Nemotron Cascade 2 30B A3B","image":false},"nvidia/nemotron-3.5-content-safety":{"name":"Nemotron 3.5 Content Safety","image":true},"google/gemini-2.5-computer-use-preview-10-2025":{"name":"Gemini 2.5 Computer Use Preview","image":true},"google/deep-research-preview-04-2026":{"name":"Gemini Deep Research Preview","image":true},"google/gemini-3.1-flash-tts-preview":{"name":"Gemini 3.1 Flash TTS Preview","image":false},"google/gemini-flash-latest":{"name":"Gemini Flash Latest","image":true},"google/gemini-embedding-2":{"name":"Gemini Embedding 2","image":true},"google/lyria-3-pro-preview":{"name":"Lyria 3 Pro Preview","image":true},"google/gemini-3.5-flash":{"name":"Gemini 3.5 Flash","image":true},"google/gemini-2.5-flash":{"name":"Gemini 2.5 Flash","image":true},"google/gemini-3.5-flash-lite":{"name":"Gemini 3.5 Flash Lite","image":true},"google/lyria-3-clip-preview":{"name":"Lyria 3 Clip Preview","image":true},"google/gemini-omni-flash-preview":{"name":"Gemini Omni Flash Preview","image":true},"google/veo-3.1-generate-preview":{"name":"Veo 3.1 Preview","image":true},"google/gemini-2.5-pro-tts":{"name":"Gemini 2.5 Pro TTS","image":false},"google/deep-research-max-preview-04-2026":{"name":"Deep Research Max Preview","image":true},"google/gemini-3-pro-image-preview":{"name":"Nano Banana Pro","image":true},"google/gemini-3.1-flash-lite-preview":{"name":"Gemini 3.1 Flash Lite Preview","image":true},"google/gemini-2.5-flash-tts":{"name":"Gemini 2.5 Flash TTS","image":false},"google/gemini-3.5-live-translate-preview":{"name":"Gemini 3.5 Live Translate Preview","image":false},"google/gemini-3-flash-preview":{"name":"Gemini 3 Flash Preview","image":true},"google/gemini-3.1-pro-preview-customtools":{"name":"Gemini 3.1 Pro Preview Custom Tools","image":true},"google/gemini-3.1-flash-lite-image":{"name":"Nano Banana 2 Lite","image":true},"google/gemini-3.1-flash-image-preview":{"name":"Nano Banana 2","image":true},"google/gemini-robotics-er-1.6-preview":{"name":"Gemini Robotics-ER 1.6 Preview","image":true},"google/gemma-4-26b-a4b-it":{"name":"Gemma 4 26B A4B IT","image":true},"google/gemini-embedding-001":{"name":"Gemini Embedding 001","image":false},"google/veo-3.1-lite-generate-preview":{"name":"Veo 3.1 Lite Preview","image":true},"google/gemini-2.0-flash":{"name":"Gemini 2.0 Flash","image":true},"google/gemini-3.6-flash":{"name":"Gemini 3.6 Flash","image":true},"google/veo-3.1-fast-generate-preview":{"name":"Veo 3.1 Fast Preview","image":true},"google/gemini-3.1-flash-lite":{"name":"Gemini 3.1 Flash Lite","image":true},"google/gemma-4-E2B-it":{"name":"Gemma 4 E2B IT","image":true},"google/gemma-4-31b-it":{"name":"Gemma 4 31B IT","image":true},"google/gemini-3.1-flash-image":{"name":"Nano Banana 2","image":true},"google/gemini-2.5-flash-image":{"name":"Nano Banana","image":true},"google/gemini-3-pro-image":{"name":"Nano Banana Pro","image":true},"google/gemini-3.1-pro-preview":{"name":"Gemini 3.1 Pro Preview","image":true},"google/gemini-flash-lite-latest":{"name":"Gemini Flash-Lite Latest","image":true},"google/gemini-2.5-pro":{"name":"Gemini 2.5 Pro","image":true},"google/gemma-4-E4B-it":{"name":"Gemma 4 E4B IT","image":true},"google/gemini-3-pro-preview":{"name":"Gemini 3 Pro Preview","image":true},"google/gemini-2.5-flash-lite":{"name":"Gemini 2.5 Flash-Lite","image":true},"google/gemini-2.0-flash-lite":{"name":"Gemini 2.0 Flash-Lite","image":true},"google/gemini-3.1-flash-live-preview":{"name":"Gemini 3.1 Flash Live Preview","image":true},"thinkingmachines/inkling":{"name":"Inkling","image":true},"sakana/fugu-ultra":{"name":"Fugu Ultra","image":true},"sakana/fugu":{"name":"Fugu","image":true},"tencent/hy3":{"name":"Hy3","image":false},"tencent/hy3-preview":{"name":"Hy3 preview","image":false},"meta/muse-spark-1.1":{"name":"Muse Spark 1.1","image":true},"meta/llama-4-scout-17b-instruct":{"name":"Llama 4 Scout 17B Instruct","image":true},"meta/llama-3.3-70b-instruct":{"name":"Llama-3.3-70B-Instruct","image":false},"meta/llama-4-maverick-17b-instruct":{"name":"Llama 4 Maverick 17B Instruct","image":true},"meituan/longcat-2.0":{"name":"LongCat-2.0","image":false},"xai/grok-4.3":{"name":"Grok 4.3","image":true},"xai/grok-4.20-0309-non-reasoning":{"name":"Grok 4.20 (Non-Reasoning)","image":true},"xai/grok-imagine-video-1.5":{"name":"Grok Imagine Video 1.5","image":true},"xai/grok-4.5":{"name":"Grok 4.5","image":true},"xai/grok-build-0.1":{"name":"Grok Build 0.1","image":true},"xai/grok-4.20-0309-reasoning":{"name":"Grok 4.20 (Reasoning)","image":true},"mistral/mistral-small-2506":{"name":"Mistral Small 3.2","image":true},"mistral/pixtral-large-latest":{"name":"Pixtral Large (latest)","image":true},"mistral/pixtral-12b":{"name":"Pixtral 12B","image":true},"mistral/magistral-medium-latest":{"name":"Magistral Medium (latest)","image":false},"mistral/mistral-large-2411":{"name":"Mistral Large 2.1","image":false},"mistral/mistral-medium-latest":{"name":"Mistral Medium (latest)","image":true},"mistral/devstral-small-2507":{"name":"Devstral Small","image":false},"mistral/mistral-medium-2505":{"name":"Mistral Medium 3","image":true},"mistral/mistral-small-latest":{"name":"Mistral Small (latest)","image":true},"mistral/mistral-nemo":{"name":"Mistral Nemo","image":false},"mistral/mistral-small-2603":{"name":"Mistral Small 4","image":true},"mistral/mistral-medium-2604":{"name":"Mistral Medium 3.5","image":true},"mistral/codestral-latest":{"name":"Codestral (latest)","image":false},"mistral/devstral-medium-latest":{"name":"Devstral 2 (latest)","image":false},"mistral/devstral-2512":{"name":"Devstral 2","image":false},"mistral/mistral-large-latest":{"name":"Mistral Large (latest)","image":true},"mistral/devstral-medium-2507":{"name":"Devstral Medium","image":false},"mistral/mistral-large-2512":{"name":"Mistral Large 3","image":true},"poolside/laguna-xs.2":{"name":"Laguna XS.2","image":false},"poolside/laguna-xs-2.1":{"name":"Laguna XS 2.1","image":false},"poolside/laguna-s-2.1":{"name":"Laguna S 2.1","image":false},"poolside/laguna-m.1":{"name":"Laguna M.1","image":false},"minimax/MiniMax-M2":{"name":"MiniMax-M2","image":false},"minimax/MiniMax-M2.7":{"name":"MiniMax-M2.7","image":false},"minimax/MiniMax-M2.1":{"name":"MiniMax-M2.1","image":false},"minimax/MiniMax-M2.5":{"name":"MiniMax-M2.5","image":false},"minimax/MiniMax-M2.5-highspeed":{"name":"MiniMax-M2.5-highspeed","image":false},"minimax/MiniMax-M2.7-highspeed":{"name":"MiniMax-M2.7-highspeed","image":false},"minimax/MiniMax-M3":{"name":"MiniMax-M3","image":true},"deepseek/deepseek-v4-flash":{"name":"DeepSeek V4 Flash","image":false},"deepseek/deepseek-v4-pro":{"name":"DeepSeek V4 Pro","image":false},"deepseek/deepseek-r1":{"name":"DeepSeek-R1","image":false},"deepseek/deepseek-chat":{"name":"DeepSeek Chat","image":false},"deepseek/deepseek-reasoner":{"name":"DeepSeek Reasoner","image":false},"deepreinforce/ornith-1.0-9b":{"name":"Ornith 1.0 9B","image":true},"deepreinforce/ornith-1.0-397b":{"name":"Ornith 1.0 397B","image":true},"deepreinforce/ornith-1.0-31b":{"name":"Ornith 1.0 31B","image":true},"deepreinforce/ornith-1.0-35b":{"name":"Ornith 1.0 35B","image":true},"alibaba/qwen3-coder-480b-a35b-instruct":{"name":"Qwen3-Coder 480B-A35B Instruct","image":false},"alibaba/qwen3.7-plus":{"name":"Qwen3.7 Plus","image":true},"alibaba/qwen3-vl-plus":{"name":"Qwen3-VL Plus","image":true},"alibaba/qwen3-32b":{"name":"Qwen3 32B","image":false},"alibaba/qwen3.8-max-preview":{"name":"Qwen3.8 Max Preview","image":true},"alibaba/qwen3.6-35b-a3b":{"name":"Qwen3.6 35B-A3B","image":true},"alibaba/qwen2-5-vl-72b-instruct":{"name":"Qwen2.5-VL 72B Instruct","image":true},"alibaba/qwen-max":{"name":"Qwen Max","image":false},"alibaba/qwen3.5-plus":{"name":"Qwen3.5 Plus","image":true},"alibaba/qwen-omni-turbo":{"name":"Qwen-Omni Turbo","image":true},"alibaba/qwen-vl-max":{"name":"Qwen-VL Max","image":true},"alibaba/qwen3-coder-30b-a3b-instruct":{"name":"Qwen3-Coder 30B-A3B Instruct","image":false},"alibaba/qwen3.5-27b":{"name":"Qwen3.5 27B","image":true},"alibaba/qwen3.7-max":{"name":"Qwen3.7 Max","image":false},"alibaba/qwen3-next-80b-a3b-thinking":{"name":"Qwen3-Next 80B-A3B (Thinking)","image":false},"alibaba/qwen3.5-9b":{"name":"Qwen3.5 9B","image":true},"alibaba/qwen3.6-27b":{"name":"Qwen3.6 27B","image":true},"alibaba/qwen3.5-35b-a3b":{"name":"Qwen3.5 35B-A3B","image":true},"alibaba/qwen3-235b-a22b":{"name":"Qwen3 235B-A22B","image":false},"alibaba/qwen3-max":{"name":"Qwen3 Max","image":false},"alibaba/qwen3-coder-plus":{"name":"Qwen3 Coder Plus","image":false},"alibaba/qwen3-coder-flash":{"name":"Qwen3 Coder Flash","image":false},"alibaba/qwen-flash":{"name":"Qwen Flash","image":false},"alibaba/qwen-plus":{"name":"Qwen Plus","image":false},"alibaba/qwen-turbo":{"name":"Qwen Turbo","image":false},"alibaba/qwq-plus":{"name":"QwQ Plus","image":false},"alibaba/qwen3.6-flash":{"name":"Qwen3.6 Flash","image":true},"alibaba/qwen3.5-397b-a17b":{"name":"Qwen3.5 397B-A17B","image":true},"alibaba/qwen3.6-max-preview":{"name":"Qwen3.6 Max Preview","image":false},"alibaba/qwen3.6-plus":{"name":"Qwen3.6 Plus","image":true},"alibaba/qwen3.5-122b-a10b":{"name":"Qwen3.5 122B-A10B","image":true},"alibaba/qwen3-next-80b-a3b-instruct":{"name":"Qwen3-Next 80B-A3B Instruct","image":false},"alibaba/qwen-vl-plus":{"name":"Qwen-VL Plus","image":true},"xiaomi/mimo-v2-omni":{"name":"MiMo-V2-Omni","image":true},"xiaomi/mimo-v2-flash":{"name":"MiMo-V2-Flash","image":false},"xiaomi/mimo-v2.5-pro-ultraspeed":{"name":"MiMo-V2.5-Pro-UltraSpeed","image":false},"xiaomi/mimo-v2-pro":{"name":"MiMo-V2-Pro","image":false},"xiaomi/mimo-v2.5":{"name":"MiMo-V2.5","image":true},"xiaomi/mimo-v2.5-pro":{"name":"MiMo-V2.5-Pro","image":false},"anthropic/claude-sonnet-4-6":{"name":"Claude Sonnet 4.6","image":true},"anthropic/claude-opus-4-0":{"name":"Claude Opus 4 (latest)","image":true},"anthropic/claude-haiku-4-5":{"name":"Claude Haiku 4.5 (latest)","image":true},"anthropic/claude-3-7-sonnet-20250219":{"name":"Claude Sonnet 3.7","image":true},"anthropic/claude-opus-4-6":{"name":"Claude Opus 4.6","image":true},"anthropic/claude-3-5-sonnet-20241022":{"name":"Claude Sonnet 3.5 v2","image":true},"anthropic/claude-fable-5":{"name":"Claude Fable 5","image":true},"anthropic/claude-opus-4-8":{"name":"Claude Opus 4.8","image":true},"anthropic/claude-sonnet-4-0":{"name":"Claude Sonnet 4 (latest)","image":true},"anthropic/claude-opus-4-1":{"name":"Claude Opus 4.1 (latest)","image":true},"anthropic/claude-sonnet-4-5":{"name":"Claude Sonnet 4.5 (latest)","image":true},"anthropic/claude-sonnet-4-5-20250929":{"name":"Claude Sonnet 4.5","image":true},"anthropic/claude-opus-4-5-20251101":{"name":"Claude Opus 4.5","image":true},"anthropic/claude-3-5-haiku-20241022":{"name":"Claude Haiku 3.5","image":true},"anthropic/claude-haiku-4-5-20251001":{"name":"Claude Haiku 4.5","image":true},"anthropic/claude-opus-4-7":{"name":"Claude Opus 4.7","image":true},"anthropic/claude-sonnet-4-20250514":{"name":"Claude Sonnet 4","image":true},"anthropic/claude-opus-4-5":{"name":"Claude Opus 4.5 (latest)","image":true},"anthropic/claude-sonnet-5":{"name":"Claude Sonnet 5","image":true},"anthropic/claude-3-haiku-20240307":{"name":"Claude Haiku 3","image":true},"anthropic/claude-opus-4-20250514":{"name":"Claude Opus 4","image":true},"anthropic/claude-opus-5":{"name":"Claude Opus 5","image":true},"anthropic/claude-opus-4-1-20250805":{"name":"Claude Opus 4.1","image":true},"perplexity/sonar":{"name":"Sonar","image":false},"perplexity/sonar-pro":{"name":"Sonar Pro","image":true},"perplexity/sonar-reasoning-pro":{"name":"Sonar Reasoning Pro","image":true},"moonshotai/kimi-k2.5":{"name":"Kimi K2.5","image":true},"moonshotai/kimi-k2-thinking-turbo":{"name":"Kimi K2 Thinking Turbo","image":false},"moonshotai/kimi-k2.7-code-highspeed":{"name":"Kimi K2.7 Code Highspeed","image":true},"moonshotai/kimi-k2.6":{"name":"Kimi K2.6","image":true},"moonshotai/kimi-k2.7-code":{"name":"Kimi K2.7 Code","image":true},"moonshotai/kimi-k2-thinking":{"name":"Kimi K2 Thinking","image":false},"moonshotai/kimi-k3":{"name":"Kimi K3","image":true},"openai/gpt-5.1-codex-mini":{"name":"GPT-5.1 Codex mini","image":true},"openai/gpt-image-2":{"name":"GPT-Image-2","image":true},"openai/gpt-5.2-pro":{"name":"GPT-5.2 Pro","image":true},"openai/gpt-5.5-pro":{"name":"GPT-5.5 Pro","image":true},"openai/gpt-4.1-mini":{"name":"GPT-4.1 mini","image":true},"openai/gpt-4o":{"name":"GPT-4o","image":true},"openai/gpt-5.4-pro":{"name":"GPT-5.4 Pro","image":true},"openai/gpt-5.6-sol":{"name":"GPT-5.6 Sol","image":true},"openai/gpt-oss-safeguard-120b":{"name":"GPT OSS Safeguard 120B","image":false},"openai/gpt-4.1-nano":{"name":"GPT-4.1 nano","image":true},"openai/gpt-oss-20b":{"name":"GPT OSS 20B","image":false},"openai/gpt-5.5-instant":{"name":"GPT-5.5 Instant","image":true},"openai/gpt-5.2-chat-latest":{"name":"GPT-5.2 Chat","image":true},"openai/o3-mini":{"name":"o3-mini","image":false},"openai/gpt-5.5":{"name":"GPT-5.5","image":true},"openai/o3-pro":{"name":"o3-pro","image":true},"openai/gpt-4o-mini":{"name":"GPT-4o mini","image":true},"openai/gpt-5":{"name":"GPT-5","image":true},"openai/o4-mini-deep-research":{"name":"o4-mini-deep-research","image":true},"openai/gpt-realtime-2.1":{"name":"GPT-Realtime-2.1","image":true},"openai/o3-deep-research":{"name":"o3-deep-research","image":true},"openai/gpt-5-codex":{"name":"GPT-5-Codex","image":true},"openai/gpt-3.5-turbo":{"name":"GPT-3.5-turbo","image":false},"openai/whisper-large-v3":{"name":"Whisper 3 Large","image":false},"openai/gpt-4o-2024-05-13":{"name":"GPT-4o (2024-05-13)","image":true},"openai/gpt-5-chat-latest":{"name":"GPT-5 Chat (latest)","image":true},"openai/gpt-4o-2024-11-20":{"name":"GPT-4o (2024-11-20)","image":true},"openai/gpt-5.4":{"name":"GPT-5.4","image":true},"openai/gpt-5.3-chat-latest":{"name":"GPT-5.3 Chat (latest)","image":true},"openai/gpt-realtime-whisper":{"name":"GPT Realtime Whisper","image":false},"openai/gpt-image-1.5":{"name":"GPT-Image-1.5","image":true},"openai/gpt-5.2-codex":{"name":"GPT-5.2 Codex","image":true},"openai/gpt-5.4-nano":{"name":"GPT-5.4 nano","image":true},"openai/gpt-5-pro":{"name":"GPT-5 Pro","image":true},"openai/gpt-5.4-mini":{"name":"GPT-5.4 mini","image":true},"openai/whisper-large-v3-turbo":{"name":"Whisper Large v3 Turbo","image":false},"openai/gpt-oss-120b":{"name":"GPT OSS 120B","image":false},"openai/o1":{"name":"o1","image":true},"openai/gpt-5.6-luna":{"name":"GPT-5.6 Luna","image":true},"openai/gpt-5.2":{"name":"GPT-5.2","image":true},"openai/gpt-5.3-codex":{"name":"GPT-5.3 Codex","image":true},"openai/gpt-5-mini":{"name":"GPT-5 Mini","image":true},"openai/o1-pro":{"name":"o1-pro","image":true},"openai/gpt-5.1":{"name":"GPT-5.1","image":true},"openai/gpt-4-turbo":{"name":"GPT-4 Turbo","image":true},"openai/gpt-4o-2024-08-06":{"name":"GPT-4o (2024-08-06)","image":true},"openai/gpt-5.1-codex":{"name":"GPT-5.1 Codex","image":true},"openai/gpt-5-nano":{"name":"GPT-5 Nano","image":true},"openai/o3":{"name":"o3","image":true},"openai/gpt-5.6-terra":{"name":"GPT-5.6 Terra","image":true},"openai/gpt-image-1":{"name":"GPT-Image-1","image":true},"openai/gpt-4.1":{"name":"GPT-4.1","image":true},"openai/o4-mini":{"name":"o4-mini","image":true},"openai/gpt-5.1-codex-max":{"name":"GPT-5.1 Codex Max","image":true},"openai/gpt-4":{"name":"GPT-4","image":false},"openai/gpt-5.1-chat-latest":{"name":"GPT-5.1 Chat","image":true},"sarvam/sarvam-30b":{"name":"Sarvam 30B","image":false},"sarvam/sarvam-105b":{"name":"Sarvam 105B","image":false},"stepfun/step-3.5-flash":{"name":"Step 3.5 Flash","image":false},"stepfun/step-3.7-flash":{"name":"Step 3.7 Flash","image":true},"stepfun/step-3.5-flash-2603":{"name":"Step 3.5 Flash 2603","image":false},"k3-256k":{"name":"K3-256k","image":true},"kimi-k3-256k":{"name":"K3-256k","image":true},"moonshot/k3-256k":{"name":"K3-256k","image":true},"kimi-for-coding":{"name":"Kimi For Coding","image":true},"moonshotai/kimi-for-coding":{"name":"Kimi For Coding","image":true},"kimi-for-coding-highspeed":{"name":"Kimi For Coding Highspeed","image":true},"moonshotai/kimi-for-coding-highspeed":{"name":"Kimi For Coding Highspeed","image":true}}
diff --git a/mcp/server.mjs b/mcp/server.mjs
new file mode 100644
index 0000000..48406c4
--- /dev/null
+++ b/mcp/server.mjs
@@ -0,0 +1,241 @@
+// kimi-eyes MCP stdio server — zero-dependency.
+// Speaks JSON-RPC 2.0 over LSP-style framing (Content-Length headers) on stdio.
+
+import fs from 'node:fs';
+import path from 'node:path';
+import {
+  VERSION,
+  getConfig,
+  requireConfigured,
+  callVLM,
+  sniffImage,
+  clipboardImageToTemp,
+  lookupModelCapability,
+} from './vision.mjs';
+
+const TOOLS = [
+  {
+    name: 'read_image',
+    description:
+      'Analyze a local image file using the configured vision API and return a text description. ' +
+      'Use when the user provides an image path or an @ file reference and your model cannot see images directly.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        path: {
+          type: 'string',
+          description: 'Absolute path to a local image file (png, jpg, jpeg, webp, gif, bmp).',
+        },
+        prompt: {
+          type: 'string',
+          description: 'Optional instruction describing what to analyze (default: describe the image).',
+        },
+      },
+      required: ['path'],
+    },
+  },
+  {
+    name: 'read_clipboard_image',
+    description:
+      'Capture the image currently in the system clipboard (e.g. a screenshot) and analyze it with the ' +
+      'configured vision API. Use when the user says they just took a screenshot or copied an image and no file path is given.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        prompt: {
+          type: 'string',
+          description: 'Optional instruction describing what to analyze (default: describe the image).',
+        },
+      },
+    },
+  },
+];
+
+// ---------------------------------------------------------------------------
+// Tools
+// ---------------------------------------------------------------------------
+
+/**
+ * When the user declares their main model (config.mainModel) and the models.dev
+ * database says it supports image input, the tools are unnecessary — the model can
+ * see images natively. Refuse with a hint instead of wasting a VLM call.
+ */
+function assertNeedsExternalVision(cfg) {
+  if (!cfg.mainModel) return;
+  const cap = lookupModelCapability(cfg.mainModel);
+  if (cap.known && cap.image) {
+    throw new Error(
+      `Model "${cfg.mainModel}" supports native image input (per the models.dev database), ` +
+        `so the kimi-eyes tools are not needed. Paste the image directly and the model will see it. ` +
+        `If you still want kimi-eyes active, remove "mainModel" from the kimi-eyes config.`
+    );
+  }
+}
+
+async function readImageTool(args) {
+  const p = args?.path;
+  if (typeof p !== 'string' || !p.trim()) {
+    throw new Error('Missing required parameter: path');
+  }
+  const cfg = getConfig();
+  assertNeedsExternalVision(cfg);
+  const abs = path.resolve(p.trim());
+  let stat;
+  try {
+    stat = fs.statSync(abs);
+  } catch {
+    throw new Error(`File not found: ${abs}`);
+  }
+  if (!stat.isFile()) throw new Error(`Not a file: ${abs}`);
+  const img = sniffImage(abs);
+  if (!img) {
+    throw new Error(
+      `Unsupported or invalid image file: ${abs} (supported: png, jpg, jpeg, webp, gif, bmp)`
+    );
+  }
+  requireConfigured(cfg);
+  return callVLM({ ...img, prompt: args?.prompt }, cfg);
+}
+
+async function readClipboardImageTool(args) {
+  const cfg = getConfig();
+  assertNeedsExternalVision(cfg);
+  requireConfigured(cfg);
+  let tmp = null;
+  try {
+    tmp = clipboardImageToTemp();
+    const img = sniffImage(tmp);
+    if (!img) throw new Error('Clipboard does not contain a valid image.');
+    return callVLM({ ...img, prompt: args?.prompt }, cfg);
+  } finally {
+    if (tmp) {
+      try {
+        fs.unlinkSync(tmp);
+      } catch {
+        // best-effort cleanup
+      }
+    }
+  }
+}
+
+async function callTool(name, args) {
+  if (name === 'read_image') return readImageTool(args);
+  if (name === 'read_clipboard_image') return readClipboardImageTool(args);
+  throw new Error(`Unknown tool: ${name}`);
+}
+
+// ---------------------------------------------------------------------------
+// JSON-RPC over stdio (MCP framing)
+// ---------------------------------------------------------------------------
+
+function send(obj) {
+  const body = Buffer.from(JSON.stringify(obj), 'utf8');
+  process.stdout.write(`Content-Length: ${body.length}\r\n\r\n`);
+  process.stdout.write(body);
+}
+
+function sendError(id, code, message) {
+  send({ jsonrpc: '2.0', id, error: { code, message } });
+}
+
+async function handleRequest(msg) {
+  const { id, method, params } = msg;
+  try {
+    switch (method) {
+      case 'initialize':
+        send({
+          jsonrpc: '2.0',
+          id,
+          result: {
+            protocolVersion: params?.protocolVersion || '2025-06-18',
+            capabilities: { tools: { listChanged: false } },
+            serverInfo: { name: 'kimi-eyes', version: VERSION },
+          },
+        });
+        return;
+      case 'ping':
+        send({ jsonrpc: '2.0', id, result: {} });
+        return;
+      case 'tools/list':
+        send({ jsonrpc: '2.0', id, result: { tools: TOOLS } });
+        return;
+      case 'tools/call': {
+        const name = params?.name;
+        const args = params?.arguments || {};
+        try {
+          const text = await callTool(name, args);
+          send({
+            jsonrpc: '2.0',
+            id,
+            result: { content: [{ type: 'text', text }] },
+          });
+        } catch (err) {
+          send({
+            jsonrpc: '2.0',
+            id,
+            result: {
+              content: [{ type: 'text', text: err.message || String(err) }],
+              isError: true,
+            },
+          });
+        }
+        return;
+      }
+      default:
+        sendError(id, -32601, `Method not found: ${method}`);
+    }
+  } catch (err) {
+    sendError(id, -32603, err.message || String(err));
+  }
+}
+
+function handleMessage(msg) {
+  if (!msg || typeof msg !== 'object' || typeof msg.method !== 'string') return;
+  const isRequest = msg.id !== undefined && msg.id !== null;
+  if (isRequest) {
+    handleRequest(msg).catch((err) => sendError(msg.id, -32603, err.message || String(err)));
+  }
+  // Notifications (e.g. notifications/initialized) are intentionally ignored.
+}
+
+// --- framing parser ---------------------------------------------------------
+
+let buf = Buffer.alloc(0);
+
+function tryParseFrame() {
+  const headerEnd = buf.indexOf('\r\n\r\n');
+  if (headerEnd === -1) return null;
+  const header = buf.subarray(0, headerEnd).toString('ascii');
+  const m = /Content-Length:\s*(\d+)/i.exec(header);
+  if (!m) return null;
+  const len = Number(m[1]);
+  const bodyStart = headerEnd + 4;
+  if (buf.length < bodyStart + len) return null;
+  const body = buf.subarray(bodyStart, bodyStart + len).toString('utf8');
+  buf = buf.subarray(bodyStart + len);
+  return body;
+}
+
+function pump() {
+  for (;;) {
+    const body = tryParseFrame();
+    if (body === null) break;
+    let msg;
+    try {
+      msg = JSON.parse(body);
+    } catch (err) {
+      console.error(`[kimi-eyes] invalid JSON frame: ${err.message}`);
+      continue;
+    }
+    handleMessage(msg);
+  }
+}
+
+process.stdin.on('data', (chunk) => {
+  buf = Buffer.concat([buf, chunk]);
+  pump();
+});
+
+process.stdin.on('end', () => process.exit(0));
+
+console.error(`[kimi-eyes] MCP server started (v${VERSION}, node ${process.version})`);
diff --git a/mcp/vision.mjs b/mcp/vision.mjs
new file mode 100644
index 0000000..c79197b
--- /dev/null
+++ b/mcp/vision.mjs
@@ -0,0 +1,344 @@
+// kimi-eyes vision core — shared by the MCP server and the setup wizard.
+// Zero-dependency. Node 18+ (native fetch, AbortSignal.timeout).
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { execFileSync } from 'node:child_process';
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+export const MODELS_DB_PATH = path.join(__dirname, 'models-db.json');
+
+export const VERSION = '1.0.0';
+
+export const EXT_MEDIA = {
+  png: 'image/png',
+  jpg: 'image/jpeg',
+  jpeg: 'image/jpeg',
+  webp: 'image/webp',
+  gif: 'image/gif',
+  bmp: 'image/bmp',
+};
+
+export const DEFAULT_PROMPT = 'Describe this image in detail.';
+
+// A minimal 1x1 PNG used for setup verification.
+export const ONE_PX_PNG_BASE64 =
+  'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==';
+
+// ---------------------------------------------------------------------------
+// Configuration
+// ---------------------------------------------------------------------------
+
+// Resolution order: environment variables (VISION_*) > config.json.
+// protocol resolves to 'openai' by default when nothing is set.
+export function getConfig(env = process.env) {
+  const home = env.KIMI_CODE_HOME || path.join(os.homedir(), '.kimi-code');
+  const cfgPath = path.join(home, 'kimi-eyes', 'config.json');
+  let file = {};
+  try {
+    file = JSON.parse(fs.readFileSync(cfgPath, 'utf8'));
+  } catch {
+    // no config file yet — fine
+  }
+  const envProtocol = (env.VISION_API_PROTOCOL || '').toLowerCase();
+  const protocol =
+    envProtocol === 'openai' || envProtocol === 'anthropic'
+      ? envProtocol
+      : file.protocol || 'openai';
+  return {
+    configPath: cfgPath,
+    protocol,
+    apiKey: env.VISION_API_KEY || file.apiKey || '',
+    baseUrl: env.VISION_API_URL || file.baseUrl || '',
+    model: env.VISION_MODEL || file.model || '',
+    mainModel: env.VISION_MAIN_MODEL || file.mainModel || '',
+  };
+}
+
+export function checkConfig(cfg) {
+  const missing = [];
+  if (!cfg.apiKey) missing.push('VISION_API_KEY');
+  if (!cfg.baseUrl) missing.push('VISION_API_URL');
+  if (!cfg.model) missing.push('VISION_MODEL');
+  return missing;
+}
+
+export function requireConfigured(cfg) {
+  const missing = checkConfig(cfg);
+  if (missing.length > 0) {
+    throw new Error(
+      `Vision API is not configured. Missing: ${missing.join(', ')}. ` +
+        `Run "node setup.mjs" in the kimi-eyes plugin directory (or set the environment variables).`
+    );
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Models.dev capability database (see scripts/sync-models.mjs)
+// ---------------------------------------------------------------------------
+
+let modelsDbCache = null; // null = not loaded yet, undefined = unavailable
+
+/** Load the slimmed models.dev database; returns null if unavailable. */
+export function loadModelsDb() {
+  if (modelsDbCache !== null) return modelsDbCache;
+  try {
+    modelsDbCache = JSON.parse(fs.readFileSync(MODELS_DB_PATH, 'utf8'));
+  } catch {
+    modelsDbCache = undefined;
+  }
+  return modelsDbCache;
+}
+
+/**
+ * Look up whether a model name supports image input, using the models.dev cache.
+ * Matching: exact id → provider/id suffix → case-insensitive display name.
+ * @returns {{known: boolean, image: boolean}}
+ */
+export function lookupModelCapability(modelName) {
+  const db = loadModelsDb();
+  if (!db) return { known: false, image: false };
+  const n = String(modelName || '').trim();
+  if (!n) return { known: false, image: false };
+  if (db[n]) return { known: true, image: !!db[n].image };
+  const lower = n.toLowerCase();
+  for (const [id, m] of Object.entries(db)) {
+    if (id.toLowerCase().endsWith(`/${lower}`)) {
+      return { known: true, image: !!m.image };
+    }
+    if (typeof m.name === 'string' && m.name.toLowerCase() === lower) {
+      return { known: true, image: !!m.image };
+    }
+  }
+  return { known: false, image: false };
+}
+
+/** Text tag used by the setup wizard: exact capability when known, keyword hint otherwise. */
+export function modelCapabilityTag(modelName, visionHintRe) {
+  const cap = lookupModelCapability(modelName);
+  if (cap.known) return cap.image ? '✓视觉' : '文本';
+  return visionHintRe && visionHintRe.test(modelName) ? '★疑似' : '';
+}
+
+// ---------------------------------------------------------------------------
+// Endpoints
+// ---------------------------------------------------------------------------
+
+export function openaiEndpoint(baseUrl) {
+  let b = String(baseUrl || '').trim().replace(/\/+$/, '');
+  if (!b) return '';
+  if (/\/chat\/completions$/i.test(b)) return b;
+  if (/\/v\d+$/i.test(b)) return `${b}/chat/completions`;
+  return `${b}/v1/chat/completions`;
+}
+
+export function anthropicEndpoint(baseUrl) {
+  let b = String(baseUrl || '').trim().replace(/\/+$/, '');
+  if (!b) return '';
+  if (/\/v1\/messages$/i.test(b)) return b;
+  if (/\/v\d+$/i.test(b)) return `${b}/messages`;
+  return `${b}/v1/messages`;
+}
+
+export function modelsEndpoint(baseUrl, protocol) {
+  let b = String(baseUrl || '').trim().replace(/\/+$/, '');
+  if (!b) return '';
+  if (/\/v\d+$/i.test(b)) return `${b}/models`;
+  return `${b}/v1/models`;
+}
+
+// ---------------------------------------------------------------------------
+// Vision API calls
+// ---------------------------------------------------------------------------
+
+const FETCH_TIMEOUT_MS = Number(process.env.VISION_FETCH_TIMEOUT_MS) || 60000;
+const MAX_TOKENS = Number(process.env.VISION_MAX_TOKENS) || 1024;
+
+/**
+ * Call the user-configured VLM with a base64 image.
+ * @param {{imageBase64:string, mediaType:string, prompt?:string}} input
+ * @param {{protocol:string, apiKey:string, baseUrl:string, model:string}} cfg
+ * @returns {Promise<string>} text description
+ */
+export async function callVLM({ imageBase64, mediaType, prompt }, cfg) {
+  requireConfigured(cfg);
+  if (cfg.protocol === 'anthropic') {
+    return callAnthropic(imageBase64, mediaType, prompt, cfg);
+  }
+  return callOpenAI(imageBase64, mediaType, prompt, cfg);
+}
+
+async function callOpenAI(imageBase64, mediaType, prompt, cfg) {
+  const endpoint = openaiEndpoint(cfg.baseUrl);
+  const body = {
+    model: cfg.model,
+    max_tokens: MAX_TOKENS,
+    messages: [
+      {
+        role: 'user',
+        content: [
+          { type: 'text', text: prompt || DEFAULT_PROMPT },
+          { type: 'image_url', image_url: { url: `data:${mediaType};base64,${imageBase64}` } },
+        ],
+      },
+    ],
+  };
+  const res = await fetch(endpoint, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      Authorization: `Bearer ${cfg.apiKey}`,
+    },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(FETCH_TIMEOUT_MS),
+  });
+  if (!res.ok) {
+    throw new Error(`Vision API HTTP ${res.status}: ${(await safeText(res)).slice(0, 300)}`);
+  }
+  const data = await res.json();
+  const content = data?.choices?.[0]?.message?.content;
+  if (typeof content === 'string' && content.trim()) return content.trim();
+  if (Array.isArray(content)) {
+    const text = content
+      .filter((p) => p && p.type === 'text' && typeof p.text === 'string')
+      .map((p) => p.text)
+      .join('\n')
+      .trim();
+    if (text) return text;
+  }
+  throw new Error(`Unexpected response from vision API: ${JSON.stringify(data).slice(0, 300)}`);
+}
+
+async function callAnthropic(imageBase64, mediaType, prompt, cfg) {
+  const endpoint = anthropicEndpoint(cfg.baseUrl);
+  const body = {
+    model: cfg.model,
+    max_tokens: MAX_TOKENS,
+    messages: [
+      {
+        role: 'user',
+        content: [
+          { type: 'image', source: { type: 'base64', media_type: mediaType, data: imageBase64 } },
+          { type: 'text', text: prompt || DEFAULT_PROMPT },
+        ],
+      },
+    ],
+  };
+  const res = await fetch(endpoint, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'x-api-key': cfg.apiKey,
+      'anthropic-version': '2023-06-01',
+    },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(FETCH_TIMEOUT_MS),
+  });
+  if (!res.ok) {
+    throw new Error(`Vision API HTTP ${res.status}: ${(await safeText(res)).slice(0, 300)}`);
+  }
+  const data = await res.json();
+  const texts = (Array.isArray(data?.content) ? data.content : [])
+    .filter((p) => p && p.type === 'text' && typeof p.text === 'string')
+    .map((p) => p.text)
+    .join('\n')
+    .trim();
+  if (texts) return texts;
+  throw new Error(`Unexpected response from vision API: ${JSON.stringify(data).slice(0, 300)}`);
+}
+
+async function safeText(res) {
+  try {
+    return await res.text();
+  } catch {
+    return '(could not read response body)';
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Image sniffing
+// ---------------------------------------------------------------------------
+
+/**
+ * Validate that a file is a supported image and return base64 + media type.
+ * @returns {{imageBase64:string, mediaType:string}|null}
+ */
+export function sniffImage(absPath) {
+  const ext = path.extname(absPath).toLowerCase().replace('.', '');
+  if (!EXT_MEDIA[ext]) return null;
+  let buf;
+  try {
+    buf = fs.readFileSync(absPath);
+  } catch {
+    return null;
+  }
+  if (buf.length === 0) return null;
+  const ascii4 = buf.toString('ascii', 0, 4);
+  let ok = false;
+  if (ext === 'png') ok = buf.subarray(0, 8).toString('hex').startsWith('89504e47');
+  else if (ext === 'jpg' || ext === 'jpeg') ok = buf.subarray(0, 3).toString('hex') === 'ffd8ff';
+  else if (ext === 'gif') ok = ascii4 === 'GIF8';
+  else if (ext === 'bmp') ok = ascii4 === 'BM';
+  else if (ext === 'webp') {
+    ok = buf.length > 12 && ascii4 === 'RIFF' && buf.toString('ascii', 8, 12) === 'WEBP';
+  }
+  if (!ok) return null;
+  return { imageBase64: buf.toString('base64'), mediaType: EXT_MEDIA[ext] };
+}
+
+// ---------------------------------------------------------------------------
+// Clipboard capture
+// ---------------------------------------------------------------------------
+
+/**
+ * Grab the image currently in the system clipboard and save it to a temp PNG.
+ * @returns {string} path to the temp PNG file
+ */
+export function clipboardImageToTemp() {
+  const out = path.join(
+    os.tmpdir(),
+    `kimi-eyes-clip-${Date.now()}-${Math.random().toString(36).slice(2, 8)}.png`
+  );
+  if (process.platform === 'win32') {
+    const script = [
+      'Add-Type -AssemblyName System.Windows.Forms;',
+      '$img = [System.Windows.Forms.Clipboard]::GetImage();',
+      "if ($img -eq $null) { Write-Error 'No image in clipboard'; exit 1 };",
+      `$img.Save('${out.replace(/'/g, "''")}', [System.Drawing.Imaging.ImageFormat]::Png)`,
+    ].join(' ');
+    execFileSync('powershell', ['-NoProfile', '-NonInteractive', '-Command', script], {
+      stdio: 'pipe',
+      timeout: 20000,
+    });
+  } else if (process.platform === 'darwin') {
+    try {
+      execFileSync('pngpaste', [out], { stdio: 'pipe', timeout: 20000 });
+    } catch (e) {
+      throw new Error('Clipboard capture failed. Install "pngpaste" (brew install pngpaste). ' + e.message);
+    }
+  } else {
+    try {
+      const png = execFileSync('wl-paste', ['-t', 'image/png', '--no-newline'], {
+        stdio: 'pipe',
+        timeout: 15000,
+      });
+      fs.writeFileSync(out, png);
+    } catch (e1) {
+      try {
+        const png = execFileSync('xclip', ['-selection', 'clipboard', '-t', 'image/png', '-o'], {
+          stdio: 'pipe',
+          timeout: 15000,
+        });
+        fs.writeFileSync(out, png);
+      } catch (e2) {
+        throw new Error(
+          'Clipboard capture failed. Need "wl-paste" (Wayland) or "xclip" (X11). ' +
+            `${e1.message}; ${e2.message}`
+        );
+      }
+    }
+  }
+  return out;
+}
diff --git a/package.json b/package.json
new file mode 100644
index 0000000..8d6cf21
--- /dev/null
+++ b/package.json
@@ -0,0 +1,40 @@
+{
+  "name": "kimi-eyes",
+  "version": "1.0.0",
+  "description": "Give non-multimodal models in Kimi Code the ability to analyze images and screenshots via a user-configured VLM (OpenAI-compatible or Anthropic protocol).",
+  "keywords": [
+    "kimi",
+    "kimi-code",
+    "vision",
+    "image",
+    "screenshot",
+    "multimodal",
+    "vlm",
+    "mcp",
+    "plugin"
+  ],
+  "license": "MIT",
+  "author": "billowliu2",
+  "type": "module",
+  "bin": {
+    "kimi-eyes": "setup.mjs"
+  },
+  "files": [
+    "mcp/",
+    "scripts/",
+    "SYSTEM.md",
+    "kimi.plugin.json",
+    "setup.mjs",
+    "README.md",
+    "README.en.md",
+    "LICENSE"
+  ],
+  "engines": {
+    "node": ">=18"
+  },
+  "scripts": {
+    "setup": "node setup.mjs",
+    "sync:models": "node scripts/sync-models.mjs",
+    "prepublishOnly": "node scripts/sync-models.mjs"
+  }
+}
diff --git a/scripts/sync-models.mjs b/scripts/sync-models.mjs
new file mode 100644
index 0000000..579ce7a
--- /dev/null
+++ b/scripts/sync-models.mjs
@@ -0,0 +1,95 @@
+#!/usr/bin/env node
+// kimi-eyes — sync the models.dev capability database into a slim local cache.
+// Run this when packaging/releasing the plugin:
+//
+//   node scripts/sync-models.mjs
+//
+// Options:
+//   --timeout <seconds>   curl timeout (default 180)
+//   --out <path>          output file (default ./mcp/models-db.json)
+//   --endpoint <url>      override the source URL (default https://models.dev/models.json)
+//
+// Proxy: the script uses the HTTPS_PROXY / https_proxy environment variable if set
+// (for users behind Clash/V2Ray etc.), falling back to a direct connection.
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { execFileSync } from 'node:child_process';
+
+const DEFAULT_ENDPOINT = 'https://models.dev/models.json';
+const DEFAULT_OUT = path.join('mcp', 'models-db.json');
+
+// Models known to support image input but missing from models.dev.
+// Maintained manually — merged into the db on every sync so re-syncs never drop them.
+const EXTRA_ENTRIES = {
+  'k3-256k': { name: 'K3-256k', image: true },
+  'kimi-k3-256k': { name: 'K3-256k', image: true },
+  'moonshot/k3-256k': { name: 'K3-256k', image: true },
+  'kimi-for-coding': { name: 'Kimi For Coding', image: true },
+  'moonshotai/kimi-for-coding': { name: 'Kimi For Coding', image: true },
+  'kimi-for-coding-highspeed': { name: 'Kimi For Coding Highspeed', image: true },
+  'moonshotai/kimi-for-coding-highspeed': { name: 'Kimi For Coding Highspeed', image: true },
+};
+
+function parseArgs(argv) {
+  const args = { timeout: 180, out: DEFAULT_OUT, endpoint: DEFAULT_ENDPOINT };
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === '--timeout') args.timeout = Number(argv[++i]) || 180;
+    else if (argv[i] === '--out') args.out = argv[++i];
+    else if (argv[i] === '--endpoint') args.endpoint = argv[++i];
+  }
+  return args;
+}
+
+function download(endpoint, outFile, timeout) {
+  const proxy = process.env.HTTPS_PROXY || process.env.https_proxy;
+  const curl = ['-sSL', '--max-time', String(timeout), '-o', outFile, endpoint];
+  const args = proxy ? ['-x', proxy, ...curl] : curl;
+  try {
+    execFileSync('curl', args, { stdio: 'inherit', timeout: (timeout + 30) * 1000 });
+  } catch (err) {
+    throw new Error(
+      `Failed to download ${endpoint}${proxy ? ` via proxy ${proxy}` : ''}. ` +
+        `If you are behind a proxy, set HTTPS_PROXY (e.g. HTTPS_PROXY=http://127.0.0.1:7897). ` +
+        `(${err.message.split('\n')[0]})`
+    );
+  }
+}
+
+const args = parseArgs(process.argv.slice(2));
+const tmpFile = path.join(os.tmpdir(), `models-dev-${Date.now()}.json`);
+
+console.log(`[sync-models] downloading ${args.endpoint} ...`);
+download(args.endpoint, tmpFile, args.timeout);
+
+const raw = JSON.parse(fs.readFileSync(tmpFile, 'utf8'));
+if (typeof raw !== 'object' || raw === null || Array.isArray(raw)) {
+  fs.rmSync(tmpFile, { force: true });
+  throw new Error(`Unexpected response shape from ${args.endpoint} (expected a flat model map).`);
+}
+
+// Slim down: keep only what the plugin needs.
+const slim = {};
+let imageCount = 0;
+for (const [id, m] of Object.entries(raw)) {
+  if (!m || typeof m !== 'object') continue;
+  const input = Array.isArray(m.modalities?.input) ? m.modalities.input : [];
+  const image = input.includes('image');
+  if (image) imageCount++;
+  slim[id] = { name: typeof m.name === 'string' ? m.name : id, image };
+}
+// Merge manual extras (they override the downloaded data on id collisions).
+for (const [id, m] of Object.entries(EXTRA_ENTRIES)) {
+  slim[id] = m;
+  if (m.image) imageCount++;
+}
+
+fs.rmSync(tmpFile, { force: true });
+fs.mkdirSync(path.dirname(args.out), { recursive: true });
+fs.writeFileSync(args.out, `${JSON.stringify(slim)}\n`, 'utf8');
+
+const kb = (fs.statSync(args.out).size / 1024).toFixed(0);
+console.log(
+  `[sync-models] wrote ${Object.keys(slim).length} models (${imageCount} support image input) to ${args.out} (${kb} KB)`
+);
diff --git a/setup.mjs b/setup.mjs
new file mode 100644
index 0000000..79c7a04
--- /dev/null
+++ b/setup.mjs
@@ -0,0 +1,241 @@
+#!/usr/bin/env node
+// kimi-eyes setup wizard — interactive one-time configuration.
+// Zero-dependency: raw-mode stdin input + native fetch.
+//
+// Flow: choose protocol → base URL → API key → pick multimodal model → verify → write config.
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { callVLM, modelsEndpoint, ONE_PX_PNG_BASE64, modelCapabilityTag } from './mcp/vision.mjs';
+
+// ---------------------------------------------------------------------------
+// Raw-mode line input (supports masked echo)
+// ---------------------------------------------------------------------------
+
+let pipeLines = null;
+function ensurePipeLines() {
+  if (!pipeLines) {
+    pipeLines = new Promise((resolve, reject) => {
+      let s = '';
+      process.stdin.setEncoding('utf8');
+      process.stdin.on('data', (c) => (s += c));
+      process.stdin.on('end', () => resolve(s.split(/\r?\n/)));
+      process.stdin.on('error', reject);
+      process.stdin.resume();
+    });
+  }
+  return pipeLines;
+}
+
+function ask(query, { hidden = false } = {}) {
+  const stdin = process.stdin;
+  if (!stdin.isTTY) {
+    // Non-TTY fallback (piped input in tests/CI): plain line read, no masking.
+    process.stdout.write(query);
+    return ensurePipeLines().then((lines) => lines.shift() ?? '');
+  }
+  process.stdout.write(query);
+  stdin.setRawMode(true);
+  stdin.resume();
+  return new Promise((resolve) => {
+    let val = '';
+    let esc = false;
+    const cleanup = () => {
+      stdin.setRawMode(false);
+      stdin.removeListener('data', onData);
+      process.stdout.write('\n');
+    };
+    const onData = (chunk) => {
+      for (const ch of chunk.toString('utf8')) {
+        if (esc) {
+          if ((ch >= 'A' && ch <= 'D') || ch === '~') esc = false;
+          continue;
+        }
+        if (ch === '\u001b') {
+          esc = true;
+          continue;
+        }
+        if (ch === '\r' || ch === '\n') {
+          cleanup();
+          resolve(val);
+          return;
+        }
+        if (ch === '\u0003') process.exit(130);
+        if (ch === '\u007f' || ch === '\b') {
+          if (val.length > 0) {
+            val = val.slice(0, -1);
+            process.stdout.write('\b \b');
+          }
+          continue;
+        }
+        val += ch;
+        process.stdout.write(hidden ? '*' : ch);
+      }
+    };
+    stdin.on('data', onData);
+  });
+}
+
+async function askRequired(query, { hidden = false, validate = null } = {}) {
+  for (;;) {
+    const v = (await ask(query, { hidden })).trim();
+    if (!v) {
+      console.log('  输入不能为空，请重试。');
+      continue;
+    }
+    if (validate) {
+      const err = validate(v);
+      if (err) {
+        console.log(`  ${err}`);
+        continue;
+      }
+    }
+    return v;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+const VISION_HINTS =
+  /(vision|visual|vlm|multimodal|4o|omni|glm-4v|qwen-vl|qwen2.*?vl|internvl|minicpm|gpt-4v|gpt-5|claude-3|claude-sonnet|claude-opus|gemini)/i;
+
+async function fetchModels(protocol, baseUrl, apiKey) {
+  const url = modelsEndpoint(baseUrl, protocol);
+  const headers =
+    protocol === 'anthropic'
+      ? { 'x-api-key': apiKey, 'anthropic-version': '2023-06-01' }
+      : { Authorization: `Bearer ${apiKey}` };
+  const res = await fetch(url, { headers, signal: AbortSignal.timeout(15000) });
+  if (!res.ok) throw new Error(`HTTP ${res.status}`);
+  const data = await res.json();
+  const list = Array.isArray(data?.data)
+    ? data.data.map((m) => m.id || m.name).filter((x) => typeof x === 'string' && x)
+    : [];
+  if (list.length === 0) throw new Error('接口未返回任何模型');
+  return [...new Set(list)];
+}
+
+async function chooseProtocol() {
+  console.log('  选择视觉 API 协议：');
+  console.log('    [1] OpenAI 兼容 (Chat Completions)');
+  console.log('    [2] Anthropic (Messages API)');
+  for (;;) {
+    const v = (await ask('  输入序号 [1/2]: ')).trim();
+    if (v === '1') return 'openai';
+    if (v === '2') return 'anthropic';
+    console.log('  无效输入，请输入 1 或 2。');
+  }
+}
+
+async function chooseModel(protocol, baseUrl, apiKey) {
+  console.log('  拉取模型列表 ...');
+  let models = null;
+  try {
+    models = await fetchModels(protocol, baseUrl, apiKey);
+  } catch (err) {
+    console.log(`  拉取模型列表失败 (${err.message})，将改为手动输入模型名。`);
+  }
+  if (models) {
+    const shown = models.slice(0, 50);
+    console.log(
+      `  可用模型 (✓视觉=支持识图 / 文本=纯文本 / ★疑似=未收录按关键词猜测): ` +
+        `${models.length > 50 ? `显示前 50 / 共 ${models.length}` : ''}`
+    );
+    shown.forEach((m, i) => {
+      const tag = modelCapabilityTag(m, VISION_HINTS);
+      console.log(`    [${String(i + 1).padStart(3)}] ${m}${tag ? ` ${tag}` : ''}`);
+    });
+    console.log('  输入序号选择模型，或直接输入自定义模型名。');
+    for (;;) {
+      const v = (await ask('  模型: ')).trim();
+      if (v === '') {
+        console.log('  输入不能为空。');
+        continue;
+      }
+      const idx = Number.parseInt(v, 10);
+      if (!Number.isNaN(idx) && idx >= 1 && idx <= shown.length) return shown[idx - 1];
+      return v; // treat as custom model name
+    }
+  }
+  return askRequired('  模型名: ');
+}
+
+async function verifyModel(cfg) {
+  console.log('  视觉验证：发送 1×1 测试图到模型 ...');
+  try {
+    const text = await callVLM(
+      { imageBase64: ONE_PX_PNG_BASE64, mediaType: 'image/png', prompt: 'Reply with the single word OK.' },
+      cfg
+    );
+    console.log(`  ✓ 通过，模型可接收图片 (返回: ${text.slice(0, 60)})`);
+    return true;
+  } catch (err) {
+    console.log(`  ✗ 失败: ${err.message.slice(0, 200)}`);
+    return false;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Main
+// ---------------------------------------------------------------------------
+
+async function main() {
+  const home = process.env.KIMI_CODE_HOME || path.join(os.homedir(), '.kimi-code');
+  const configPath = path.join(home, 'kimi-eyes', 'config.json');
+
+  console.log('=== Kimi Eyes 配置向导 ===');
+  console.log(`配置将写入: ${configPath}\n`);
+
+  const protocol = await chooseProtocol();
+  const baseUrl = await askRequired('  视觉 API BaseUrl (如 https://api.openai.com/v1): ', {
+    validate: (v) => (/^https?:\/\//i.test(v) ? null : 'BaseUrl 必须以 http:// 或 https:// 开头。'),
+  });
+  const apiKey = await askRequired('  API Key: ', { hidden: true });
+  const model = await chooseModel(protocol, baseUrl, apiKey);
+
+  const cfg = { protocol, baseUrl, apiKey, model };
+
+  console.log('');
+  for (let attempt = 0; attempt < 3; attempt++) {
+    if (await verifyModel(cfg)) break;
+    if (attempt === 2) {
+      console.log('  连续验证失败，跳过验证（配置仍将写入）。');
+      break;
+    }
+    const act = (await ask('  [r] 重试  [c] 换模型  [s] 跳过验证: ')).trim().toLowerCase();
+    if (act === 'c') {
+      cfg.model = await chooseModel(protocol, baseUrl, apiKey);
+      continue;
+    }
+    if (act === 's') break;
+    // 'r' or anything else → retry the loop
+  }
+
+  const mainModel = (await ask('  当前主模型名（可选，用于触发判断；留空跳过）: ')).trim();
+  if (mainModel) cfg.mainModel = mainModel;
+
+  const dir = path.dirname(configPath);
+  fs.mkdirSync(dir, { recursive: true });
+  fs.writeFileSync(configPath, `${JSON.stringify(cfg, null, 2)}\n`, 'utf8');
+  if (process.platform !== 'win32') fs.chmodSync(configPath, 0o600);
+
+  console.log('');
+  console.log(`✓ 配置已写入: ${configPath}`);
+  console.log(`  protocol: ${cfg.protocol}`);
+  console.log(`  baseUrl : ${cfg.baseUrl}`);
+  console.log(`  model   : ${cfg.model}`);
+  if (cfg.mainModel) console.log(`  mainModel: ${cfg.mainModel} (触发判断用)`);
+  console.log('');
+  console.log('下一步：');
+  console.log(`  1. 安装插件: /plugins install ${process.cwd()}`);
+  console.log('  2. 启用插件: /reload');
+  console.log('  3. 开始使用: 非多模态模型下 @图片路径 或 截图后提问');
+}
+
+main().catch((err) => {
+  console.error(`\n[setup] 出错: ${err.message}`);
+  process.exit(1);
+});