Intercept, trace, and swap LLMs without leaving the terminal

What it does Lemmy is a TypeScript monorepo that pairs a multi-provider LLM client with a terminal UI framework. The core library normalizes Anthropic, OpenAI, and Google Gemini behind one interface, handles conversation context manually, and requires you to execute tools yourself rather than letting the model run unsupervised. The TUI package renders terminal interfaces with differential updates, and the apps folder includes utilities like claude-bridge to reroute Claude Code traffic to other providers and claude-trace to visualize those conversations.

The interesting bit The project keeps humans in the loop by design: tool calls are exposed for interception instead of being hidden inside a black-box loop. claude-bridge effectively decouples Claude Code’s IDE integration from Anthropic’s API, while claude-trace turns opaque agent conversations into inspectable artifacts.

Key highlights

• Unified LLM client for Anthropic, OpenAI, and Google with Zod-typed tool definitions
• Manual tool execution with interception hooks for controlled workflows
• Terminal UI framework with differential rendering and composable components
• claude-bridge intercepts Claude Code requests to use other LLM providers
• claude-trace visualizes Claude Code conversations; covered by Simon Willison

Caveats

• lemmy-chat is explicitly marked work in progress
• Several apps (red-teaming, snap-happy) are narrow examples or single-purpose utilities
• Package-level documentation lives in subfolder READMEs; the top-level README is an overview

Verdict Developers building terminal-first AI tools or trying to audit and redirect Claude Code traffic will find the most value here; those looking for a fully autonomous agent framework should look elsewhere.