Ouroboros is a local-first runtime layer for AI coding work that sits between a person and a coding agent. The coding agent writes the code, and Ouroboros pins the specification before the run and verifies the result after it. Every action becomes a Seed-bound, ledger-recorded, replayable event, whichever model executes it.
Work moves through a loop: a Socratic interview exposes hidden assumptions and scores the remaining ambiguity, the answers crystallize into an immutable specification (the Seed) with acceptance criteria, ontology, and constraints, execution follows a Double Diamond decomposition, and a three-stage gate evaluates the result (mechanical checks, then semantic review, then multi-model consensus). Evaluation output feeds the next generation's Seed, and the loop stops when ontology similarity reaches 0.95 or after 30 generations. An ambiguity gate blocks code generation until the interview scores 0.2 or lower, unless forced. The grading command and expected result stay out of the success contract handed to the agent.
Ouroboros runs as ooo skills inside a coding agent session, as an MCP server, or as a standalone CLI. Supported runtimes include Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Hermes, Gemini, Kiro CLI, Pi CLI, OMP CLI, Zcode, Goose, GJC, Antigravity CLI, and Grok Build CLI, and the same Seed runs on any of them. It can also send the interview, Seed, and QA stages to DeepSeek models through DeepSeek Harness, or run inside DeepSeek Harness as a plugin. A LiteLLM adapter covers other providers.
The core is one layer of a three-part stack: a separate plugins repository composes core primitives into domain workflows (PR operations, Jira sync, incidents, releases) with scoped permissions, and a separate terminal client, Ourocode, runs ooo workflows across several coding CLIs in one session.
Features
- Socratic interview: a questions-only agent surfaces assumptions and reports an ambiguity score before any code is written
- Immutable Seed: a locked specification with acceptance criteria, ontology, and constraints that each run executes against
- Three-stage evaluation: mechanical checks, semantic review, and multi-model consensus gate each result
- Evolutionary loop: Wonder and Reflect steps feed evaluation output into the next generation until the ontology converges
- Ralph: runs the evolutionary loop persistently across sessions, rebuilding lineage from an event store after a restart
- Specialist agents: 21 on-demand agents, including Ontologist, Seed Architect, Evaluator, Contrarian, Hacker, Simplifier, Researcher, and Architect
- PAL Router: three cost tiers (frugal, standard, frontier) with escalation on failure and downgrade on success
- Drift and stagnation detection: weighted goal, constraint, and ontology drift plus detection of spinning, oscillation, and diminishing returns
- Multi-runtime adapter: one workflow spec across Claude Code, Codex CLI, OpenCode, Hermes, Gemini, Goose, Kiro, Copilot, Pi, and others
- Team publishing: turns a Seed into GitHub epic and task issues, and triages maintainer issues and PRs
