Ouroboros

What it does

Ouroboros describes itself as an "Agent OS" for replayable AI coding workflows. It starts with a Socratic interview that exposes hidden assumptions, crystallises the answers into an immutable specification called a Seed, then hands that Seed to a coding agent runtime and checks the result through a three-stage evaluation gate. Every action is recorded in an event ledger so runs can be replayed and inspected.

The ouroboros-ai Python package includes an MCP server, started with `ouroboros mcp serve`, which exposes tools for interviewing, generating and executing Seeds, polling background jobs, running QA verdicts and querying session, project and event status. Its installer registers the server with the coding agents it finds, including Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Gemini, Kiro CLI and Goose; inside those hosts it is driven with `ooo` commands such as `ooo interview`, `ooo run` and `ooo evaluate`.

It needs Python 3.12 or later and one supported agent runtime or LLM backend, which brings its own sign-in or API key (for example ANTHROPIC_API_KEY or OPENAI_API_KEY). The docs say to use version 0.51.1 or later when installing it as an MCP server. The server speaks stdio by default and can also serve SSE or streamable HTTP.

What it needs