Pi T3chat

pi-t3chat running in Pi — 50+ frontier models via t3.chat subscription

pi-t3chat

t3.chat models in Pi — 50+ frontier AI models via your t3.chat subscription

Claude Fable-5 · GPT-5.6 · Gemini 3.5 · Grok 4.3 · DeepSeek V4 · Llama 4 · Qwen 3 · 50+ more

Overview

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.

How It Works

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.

Features

• 50+ frontier AI models — Claude, GPT, Gemini, Grok, DeepSeek, Llama, Qwen, and more through a single t3.chat subscription
• Dynamic model discovery — scrapes t3.chat's JS bundles to fetch model definitions; zero hardcoding
• TLS fingerprint impersonation — wreq-js with Chrome 142 emulation bypasses t3.chat's bot detection
• Text-based tool calling — injects tool definitions as text; model emits tool: fenced blocks; proxy converts to OpenAI tool_calls
• MCP wrapper tools — mcp__ prefixed tools are grouped into MCP servers with list_mcps / list_mcp_tools / call_mcp discovery protocol
• False refusal correction — auto-retries when the model claims it "can't access tools" while tools are available
• Credit tracking — real-time balance, usage percentages, and subscription tier via tRPC endpoints
• Cookie-based auth — no API keys; uses your existing t3.chat browser session

Tool Calling

t3.chat models don't all support OpenAI-native function calling. This extension implements a text-based tool calling protocol:

1. Injection — OpenAI tool definitions are injected into the system prompt as text instructions
2. Emission — The model emits tool:<name> fenced code blocks in its response
3. Parsing — The proxy parses these blocks and converts them to OpenAI tool_calls format
4. False refusal correction — If the model claims it "can't access tools" while tools are available, the proxy automatically retries with a correction prompt (up to 2 retries)

Example 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.

MCP Wrapper Tools

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 groups
• list_mcp_tools — List tools within one MCP group
• call_mcp — Execute a specific MCP tool by group + tool name

The model discovers tools on-demand through a multi-round loop (max 6 rounds), reducing prompt size dramatically when many MCP tools are available.

Installation

Git (recommended)

pi extension add https://github.com/vibheksoni/pi-t3chat.git

Local dev

git clone https://github.com/vibheksoni/pi-t3chat.git
cd pi-t3chat
pi extension add .

Setup

1. Sign in to t3.chat in your browser
2. Open DevTools → Application → Cookies → t3.chat
3. Run /login t3chat in Pi
4. Paste your full Cookie header string
5. Paste your convex-session-id value
6. Select a model with /model t3chat/<model-id>

Commands

Models

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.

Notable Models

TLS Impersonation

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.

File Structure

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

Endpoints Used

License

MIT