Centauri is a harness-preserving Agent Development Environment for Claude Code, Codex, Droid, OpenCode, Aider, Vibe, and the CLI coding agents you already use.
It does not replace your agent loop. It launches your selected local harness, keeps its native behavior intact, and gives you one Git-native workspace around it.
Different coding agents are not just different model endpoints. They carry prompting strategies, tool policies, permission models, context handling, and model-specific assumptions. If you trust Claude Code for Opus, Codex for OpenAI models, OpenCode for GLM, or Droid for a particular workflow, you should not have to give that up to get a better workspace.
Centauri pairs agent sessions with source control so you can prompt, inspect, review, stage, commit, and push from one local loop.
Run different coding agents side-by-side while preserving each harness's native behavior and trust model.
Use chat-capable CLIs in a calmer conversational view while the actual harness still runs locally in the repo.
Drop into raw TUI mode when a harness works best as the toolmaker intended.
Inspect diffs, stage and unstage files, discard intentionally, generate commit messages, commit, and push.
Repositories stay on your filesystem. Git uses your system executable. GitHub uses your existing credentials.
Claude Code, Codex, Droid, pi, Mistral Vibe, OpenCode, Aider, Gemini CLI, Cursor Agent, Amp, Hermes, and more.
Centauri is early and moving fast. The core local agent + Git loop is working, while packaging, session persistence, and harness-specific polish are still evolving.
git clone https://github.com/unmodeled-tyler/centauri.git
cd centauri
npm install
npm run dev
A real browser runtime for web agents: persistent sessions, MCP control, page-change tracking, and visible human supervision.
A local ADE for coding agents: preserve your chosen harnesses, compare their work, and ship through a Git-native workspace.