# AgentRouter

> One key for models, APIs, MCP servers and Skills. Pay per call; failed calls are free.

## Setup
1. Ask the user to create an API key at http://localhost:3000/console/keys (never ask them to paste it into chat; store it as the env var AR_KEY).
2. Send `Authorization: Bearer $AR_KEY` on every request.
3. Best: add the MCP server http://localhost:3000/api/mcp (same header), or install the skill at http://localhost:3000/skill.md.

## Endpoints
- Discover: GET http://localhost:3000/api/v1/capabilities?q=<task>&kind=model|api|mcp|skill  (search the catalog; do not guess ids)
- Inspect: GET http://localhost:3000/api/v1/capabilities/<id>  (parameters and price)
- Quote: POST http://localhost:3000/api/v1/quote {"capability":"<id>"}
- Run: POST http://localhost:3000/api/v1/run {"capability":"<id>","input":{...},"policy":"balanced|cheapest|fastest|quality|priority"}
- Models (OpenAI-compatible): POST http://localhost:3000/api/v1/chat/completions ; list: GET http://localhost:3000/api/v1/models
- Data APIs as REST: GET|POST http://localhost:3000/api/data/<capability id without "api/">
- Third-party MCP tools (Stripe, Notion, 飞书 …): the user connects once at http://localhost:3000/console/connections, then run them like any capability.

A 202 response with status "pending_approval" means a team admin must approve the call.

## Catalog
- Models (6): 文本与推理, 向量, 图像, 视频
- APIs (8): 社媒, 搜索与抽取, 数据库, 电商, 金融
- MCP (3): 办公协作, 开发, 数据库
- Skills (71): 研究与情报, 内容创作, 增长与 SEO, 投研与金融, 学术写作, 法律, 行业专业, 入门, 编程

## Featured
- api/web.extract — 网页正文抽取: 任意 URL 转为干净 Markdown，支持 JS 渲染。
- api/x.search — X 帖子搜索: 按关键词、账号、时间范围检索公开帖子。
- mcp/github — GitHub: 42 个工具：仓库、Issue、PR、Actions。OAuth 由平台代管。
- model/claude-opus-5-5 — Claude Opus 5.5: 复杂推理、长上下文编程与多步 Agent 任务的旗舰模型。
- skill/pkg/agentrouter — AgentRouter（官方）: 让 Agent 学会使用 AgentRouter：搜索、查看、报价、运行；模型、数据接口、MCP 工具与 Skills 一把钥匙。建议第一个安装。
- skill/social-radar — 竞品社媒雷达: 每日监控 X、Reddit、小红书的品牌声量与情绪，输出一页简报。
- skill/pr-review — PR 代码审查: 读取 diff 与 CI 日志，按严重度给出可执行的修改建议。