Kiln

The open source software factory. AI agents plan, write, review and verify your code. A self-hosted alternative to Warp Factories. · kiln.fishu.workers.dev

CI/CD is the assembly line that builds, tests and ships code people wrote. A software factory extends that line to the whole development lifecycle: AI agents plan the change, write it, review it and prove it works, and your CI/CD still runs on every pull request. Issues, Slack threads, tracker tickets and schedules go in. Each one runs through a pipeline you define: triage, spec, build in a sandbox, tests, a preview, review. A reviewed pull request comes out. People approve the spec, accept the preview and merge. Scorers grade every run, and repeated failures turn into pull requests that improve the factory itself.

• Bring your own key. No Kiln-hosted model and no markup. Plug in what you already pay for:
  • Claude: a Claude subscription (claude setup-token) or an Anthropic API key
  • Codex: an OpenAI API key
  • AWS Bedrock: a Bedrock API key, or AWS access keys and a region, for Claude Code on Bedrock
  • Others: OpenCode and Pi with OpenRouter or Gemini keys
• Self-hosted. One process, a data folder and Docker to start; Docker Compose or Helm (gVisor sandboxes) for teams. Your code and keys stay on your infrastructure.
• Pipelines as code. A tree of factory.yaml, agents/, automations/, pipelines/, scorers/, skills/ … stored in a GitHub repo you own and reviewed like any other change. Merges apply automatically.
• People stay in charge. Agents push only kiln/* branches with short-lived tokens and can never approve or merge.

Build your factory with your favorite AI

kiln architect

Then ask your coding agent to build it: /kiln-architect in Claude Code, or "Follow .kiln/architect.md" in Codex, Cursor, Copilot, Gemini CLI or anything else. The agent profiles your repo (kiln scan), reads your merged pull requests, and with your OK your tracker and Slack. It writes a factory that fits how your team already works, backtests it on your last merged PRs (kiln bench from-prs), and explains every choice with its evidence in FACTORY.md. Nothing is applied until you say yes. Guide

Quickstart

pnpm install
pnpm dev                 # API on :4000 (with demo data), web app on :3000

Open http://localhost:3000. Docker must be running: every agent run starts in a Docker sandbox, and pnpm dev starts the sandbox egress proxy for you (docker compose up -d --build egress-proxy). Out of the box everything else is local: a git-backed forge, an in-process tracker and chat, and scripted agents, so you can see the whole lifecycle without accounts. To use real models, set KILN_AGENTS=real and one of CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_API_KEY, OPENAI_API_KEY or OPENROUTER_API_KEY. See Self-hosting for Docker Compose, Helm and connecting GitHub, Slack, Linear and Jira.

pnpm kiln init my-factory --template sdlc    # or single-repo, po-feature, incident
pnpm kiln validate my-factory
pnpm kiln apply my-factory                   # with KILN_URL and KILN_TOKEN set
pnpm kiln send "Fix the typo in the README" --factory <alias>

How it works

A work item moves through triaging → planning → awaiting_spec → building → awaiting_preview → reviewing → handed_off → complete. The foreman picks the shortest path (small, clear changes skip the spec), or a pipeline file fixes the graph. Each run happens in a sandbox; every shell command passes a guardrail; evidence (tests, a video per acceptance criterion, a preview URL) is attached to the PR. Scorers grade runs, and self-improvement turns repeated failures into definition PRs that are replayed against the failing runs before a person reviews them.

Read more in the docs (source in docs/site, also served at /docs in the web app). Design decisions are in docs/adr.

Repository

apps/api        REST API, webhooks, MCP endpoint, timers (Hono)
apps/mcp        Kiln MCP server (streamable HTTP, scoped tokens)
apps/web        Web app (Next.js + shadcn/ui)
packages/       schema, definitions, decide, sandbox, harness, evidence,
                integrations, auth, store, events, engine, sdk-ts, cli
factories/      starter factory templates
docs/           docs site, ADRs, threat model, ideas, screenshots
deploy/         Helm chart and the sandbox egress proxy
e2e/            end-to-end tests on a git-backed fake forge

Development

pnpm typecheck && pnpm test && pnpm lint && pnpm build
pnpm docs:gen      # regenerate CLI/API/MCP/definition reference pages

Contributions follow CLAUDE.md: one task per branch, tests first, a person merges. MIT licensed.