Agent Chassis

A modular, scalable foundation for building AI agents with FastAPI, OpenAI, and MCP.

You can view the full documentation here with Mintlify.

Why it’s useful

• Hybrid tools (local Python + MCP) with OpenAI-compatible schemas.
• Streaming and non-streaming agent loop with tool calling.
• Optional persistence (Redis/Postgres) and auth (API key / JWT).

Quick start (bite-sized)

1. Install deps

uv sync

2. Configure env

cp .env.example .env  # set OPENAI_API_KEY, CHASSIS_API_KEY (optional)

3. Run the API

uv run uvicorn app.main:app --reload

API lives at http://localhost:8000 (see Mintlify for endpoints and tool setup).

Basic usage

• Add local tools in app/services/local_tools.py with @local_registry.register.
• Add MCP servers in mcp_config.json (stdio/SSE/streamable-http).
• Restrict allowed tools per request via allowed_tools.

Testing

uv run pytest

What is Agent Chassis?

Agent Chassis is a modular, asynchronous foundation for building AI agents with FastAPI, OpenAI, and MCP (Model Context Protocol). It serves as a robust "scaffolding" for agentic applications, handling the heavy lifting of infrastructure, protocol management, and tool execution so developers can focus on agent logic.

Core Components

1. Agent Service (app/services/agent_service.py): The central brain that orchestrates a multi-turn Plan-Act-Observe loop with streaming support and robust error handling.

2. MCP Manager (app/services/mcp_manager.py): Manages connections to Model Context Protocol servers via stdio or Server-Sent Events (SSE) protocols.

3. Local Tool Registry (app/services/local_tools.py): A decorator-based system for exposing local Python functions as tools to the agent.

4. Tool Translator (app/services/tool_translator.py): Universal adapter that converts MCP Tools and Python functions to OpenAI JSON Schema format.

5. Security & Configuration (app/core/): Manages API key authentication, settings via pydantic-settings, and environment configuration.

Key Features

• Hybrid Tooling: Combine remote MCP tools with local Python functions seamlessly
• Streaming & Non-Streaming: Choose between real-time token-by-token feedback or blocking responses
• Session Persistence: Optional Redis + PostgreSQL storage for conversation history
• Authentication: Support for API key and JWT-based auth with Google OAuth
• OpenAI Compatible: Works with OpenAI API and compatible providers

Next Steps

• View the full documentation at agent-chassis.techlitnow.com
• Check out AGENTS.md for a more in-depth codebase walkthrough
• See CLAUDE.md for detailed context for AI assistants
• Explore the /docs directory for comprehensive guides on configuration, tool development, and MCP integration