paradigmxyz/centaur

What it solves

Centaur provides a self-hosted platform for teams to share AI agents instead of relying on fragmented, local setups. It allows teams to interact with agents via Slack or an API, giving those agents a secure, isolated environment to execute real work—such as investigating CI failures or running operational workflows—without exposing raw API keys to the agent's environment.

How it works

Centaur uses a Rust-based control plane to manage agent sessions and durable state. When a user interacts with an agent (via Slack or API), Centaur assigns a Kubernetes sandbox containing a shell, workspace, and development tools (Python, Node.js, etc.). The agent runs within this sandbox using a chosen harness (like Claude Code or Codex). To interact with the outside world, the sandbox uses a proxy that injects credentials only during outbound requests, ensuring the agent never sees the raw secrets.

Who it’s for

It is designed for engineering and operations teams who need standardized, shared AI agents that can safely access internal repositories, run tests, and execute multi-step workflows across a team.

Highlights

• Slack-native integration: Manage agent conversations and track progress directly within Slack threads.
• Isolated execution: Every conversation runs in its own Kubernetes sandbox with a dedicated shell and workspace.
• Agent agnostic: Supports various CLI-based agent harnesses and custom Python tool plugins.
• Durable workflows: Supports long-running jobs that can sleep, resume, and survive service restarts.
• Secure credential handling: Uses a proxy to inject secrets into outbound requests, preventing raw API keys from entering the sandbox.