pi-t3chat running in Pi — 50+ frontier models via t3.chat subscription
接入你的智能体
从发布者处安装。不是 PlugMCP 端点。
远程 URL 还是本机命令? 两种安装方式的区别
来自发布者
pi-t3chat running in Pi — 50+ frontier models via t3.chat subscription
pi-t3chat t3.chat extension for Pi coding agent. 50+ frontier AI models (Claude Fable-5, GPT-5.6, Gemini 3.5, Grok 4.3, DeepSeek V4, Llama 4, Qwen 3) via t3.chat subscription. TLS impersonation, dynamic model discovery, tool calling, MCP wrapper tools.
Claude Fable-5 · GPT-5.6 · Gemini 3.5 · Grok 4.3 · DeepSeek V4 · Llama 4 · Qwen 3 · 50+ more
pi-t3chat is a Pi coding agent extension that bridges t3.chat's multi-model API into Pi's OpenAI-compatible provider interface. It spins up a local proxy that translates OpenAI Chat Completions calls to t3.chat's SSE streaming format — with full TLS fingerprint impersonation, dynamic model discovery, text-based tool calling, and MCP wrapper tool discovery.
Pi → Local Proxy (127.0.0.1:42101) → t3.chat API (https://t3.chat/api/chat)
OpenAI-compatible HTTP SSE streaming with TLS impersonation
The extension spins up a local HTTP proxy that translates OpenAI Chat Completions API calls to t3.chat's SSE format. All requests to t3.chat use wreq-js for TLS fingerprint impersonation (Chrome 142) — standard fetch() gets blocked by t3.chat's bot detection.
tool: fenced blocks; proxy converts to OpenAI tool_callsmcp__ prefixed tools are grouped into MCP servers with list_mcps / list_mcp_tools / call_mcp discovery protocolt3.chat models don't all support OpenAI-native function calling. This extension implements a text-based tool calling protocol:
tool:<name> fenced code blocks in its responsetool_calls formatExample tool block emitted by the model:
```tool:read_file
{"path": "/src/index.ts"}
```
For models that support native tool calling via the SSE stream, the proxy also handles structured tool_calls deltas.
Tools with mcp__ prefixes (e.g. mcp__exa__web_search) are automatically grouped into MCP servers. Instead of injecting all MCP tool definitions into the system prompt, the proxy injects three wrapper tools:
list_mcps — List all available MCP groupslist_mcp_tools — List tools within one MCP groupcall_mcp — Execute a specific MCP tool by group + tool nameThe model discovers tools on-demand through a multi-round loop (max 6 rounds), reducing prompt size dramatically when many MCP tools are available.
pi extension add https://github.com/vibheksoni/pi-t3chat.git
git clone https://github.com/vibheksoni/pi-t3chat.git
cd pi-t3chat
pi extension add .
t3.chat/login t3chat in Piconvex-session-id value/model t3chat/<model-id>Models are fetched dynamically by scraping t3.chat's JavaScript bundles — zero hardcoding. The catalog is cached for 10 minutes and refreshed on demand.
Run t3chat-refresh after new models are added to t3.chat.
t3.chat blocks requests with non-browser TLS fingerprints. This extension uses wreq-js — a Node.js/TypeScript HTTP client powered by native Rust wreq bindings — to impersonate Chrome 142's TLS/JA3/HTTP2 fingerprint on every request.
pi-t3chat/
├── index.ts — Pi extension entry point, provider registration, commands
├── proxy.ts — OpenAI-compatible HTTP proxy → t3.chat SSE
├── chat.ts — t3.chat SSE streaming via wreq-js
├── sse.ts — SSE stream parser for t3.chat events
├── tools.ts — Text-based tool calling protocol (injection, parsing, correction)
├── mcp.ts — MCP wrapper tools (list_mcps / list_mcp_tools / call_mcp)
├── catalog.ts — Model discovery by scraping t3.chat JS bundles
├── models.ts — Model resolution (name → ID lookup)
├── usage.ts — Credit tracking via tRPC endpoints
├── auth.ts — Cookie-based credential storage
├── config.ts — Chat config options
└── package.json — Package metadata
MIT
pi-t3chat t3.chat extension for Pi coding agent. 50+ frontier AI models (Claude Fable-5, GPT-5.6, Gemini 3.5, Grok 4.3, DeepSeek V4, Llama 4, Qwen 3) via t3.chat subscription. TLS impersonation, dyna
概览
pi-t3chat t3.chat extension for Pi coding agent. 50+ frontier AI models (Claude Fable-5, GPT-5.6, Gemini 3.5, Grok 4.3, DeepSeek V4, Llama 4, Qwen 3) via t3.chat subscription. TLS impersonation, dynamic model discovery, tool calling, MCP wrapper tools.
工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式,而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
Pi T3chat 通过 stdio 在本地运行。
上次检查时间:2026-10-28。 pi-t3chat t3.chat extension for Pi coding agent. 50+ frontier AI models (Claude Fable-5, GPT-5.6, Gemini 3.5, Grok 4.3, DeepSeek V4, Llama 4, Qwen 3) via t3.chat subscription. TLS impersonation, dynamic model discovery, tool calling, MCP wrapper tools.