An open-source autonomous-SDLC governance layer for AI agent teams — built on Multica. Give Guild a product idea; it produces a staged delivery plan, gates each stage on your approval, dispatches contracted work to coding agents running on your self-hosted Multica instance, validates every handoff against machine-checkable contracts — never an agent's self-report — and enforces a spend budget with a kill-switch.
Guild deliberately does not rebuild the agent platform. Multica ships the board, 14+ CLI runtimes, and skills; Guild is the discipline on top: stages, gates, contracts, budgets. Those are precisely the gaps verified absent in Multica (flat gate-free issues — multica#815/#1943; trusted self-reports — multica#1579; cost recorded but never enforced — see docs/research/).
Non-commercial by design: built for personal self-hosting, published open source (Apache-2.0 for Guild's own code). Multica itself is source-available — hosting it for third parties requires their commercial license, which Guild's scope deliberately avoids.
| Doc | Contents |
|---|---|
| CLAUDE.md | Development guidelines: hexagonal/DDD layering, TDD/BDD discipline, conventions (mirrored as AGENTS.md) |
| docs/PRODUCT.md | What Guild does, flows, MVP cut |
| docs/OVERVIEW.md | The map: every component and every data/event flow, with diagrams |
| docs/ARCHITECTURE.md | Decision records (D1 onward, with honest superseded/retained statuses), engagement lifecycle, deployment topology |
| docs/AGENT-MANUAL.md | The agent-facing manual: how-tos for driving Guild end to end, each backed by a named test or check |
| docs/ROADMAP.md | Milestones M0–M4 with acceptance criteria |
| docs/VALIDATION-2026-07-29.md | External validation of the original decisions, with sources |
| docs/research/external-reviews-2026-07-30.md | Four independent non-Anthropic model reviews — all blockers dispositioned: disposition table |
| docs/research/ | Full evidence trail: multica comparison, source-level investigation, review responses |
Two supported deployment targets, lowest barrier first — full details in deploy/README.md:
- Docker Compose — any machine with Docker, no Kubernetes; the tier every functional milestone (M1–M3, through the feature-complete product) ships on and exercises continuously. Run the Tier 1 quickstart →
- Any Kubernetes (optional, last — pursued only if needed, at M4) — vanilla manifests: any StorageClass or external databases (dual-mode), plain
kubectlsecrets, hardening (NetworkPolicies, gVisor) recommended where your infra supports it, never assumed.
(The author's hardened cluster is documented as a personal, non-normative runbook — required for nobody.)
packages/
shared/ published language: stages, HandoffContract, ExecutionSubstrate + ModelGateway ports
orchestrator/ the Guild conductor: contract validator (docker-run sandbox driver), LiteLLM gateway adapter, M1 smoke feature
substrate-multica/ ExecutionSubstrate adapter over Multica's REST/WS API (anti-corruption layer)
substrate-conformance/ reusable ExecutionSubstrate port contract suite — mandatory-green on every Multica pin bump
deploy/ deployment options: compose (shipped, M1) / any-K8s (optional generic manifests at M4)
docker/daemon/ custom Multica daemon image (delivered at M1) — pinned CLIs + git + headless login + LiteLLM routing
docs/ product, architecture, roadmap, validation evidence, research
(as of 2026-08-04) M1, M2a, and M2b functionally complete — the full application runs end to end on the live stack. M1 (accepted 2026-07-31) proved the substrate: the Tier 1 compose stack with the doctor diagnostic, the pinned daemon image (OpenCode default; Claude Code was in scope then and removed by the 2026-08-04 D9 amendment), the ExecutionSubstrate port with its Multica adapter and conformance suite (the standing pin-bump gate), and the first SHA-pinned contract validation through the sandboxed docker run driver. M2a (accepted 2026-08-03) ran one real engagement through the whole governed lifecycle including a genuine bounce and a conductor-kill recovery. M2b (accepted 2026-08-04, pnpm smoke:m2b green twice) is the whole product loop: an operator idea posted as a board ticket became a deterministic five-stage plan, was amended and re-gated live, and was delivered by the fixed four-role team through five sequential operator-approved gates — a repo whose tests pass in a fresh clone, at a measured 77–111¢ against a $3.00 plan — while an induced overspend halted the pipeline cleanly (work cancelled, dispatch locked, explanation on the ticket) and the decision trail audited to zero un-contracted advances. The conductor ships as a compose service with guild init / guild demo / guild kill verbs, and the fresh-machine bar ran for real: down -v → quickstart verbatim → first accepted stage in 4.5 minutes, full delivery in 17.5 (bar: ≤30). v0.2.0 is the current release — M3 team evolution, the feature-complete product: API-backed dynamic hiring (D16), archive-only retirement, role memory across engagements, the template catalog (standard/quick-fix), and run-pinned project rules — accepted live 2026-08-12 (a tester hired because the plan demanded it, contracted work delivered through gates, retired on completion; a second run's tester received the first run's role memory; hire/retire rows queryable in decisions). v0.1.1 was the prior security patch (#17, advisory GHSA-7pg8-mmpv-r6pc), a breaking upgrade for v0.1.0 installs (see deploy/compose/README.md). Kubernetes stays an optional last milestone (M4). Every design claim above traces to cited evidence in docs/.
Requires Node ≥ 22 and pnpm 10 (note: corepack is no longer bundled with Node ≥ 25 — install pnpm directly):
npm install -g pnpm@10
pnpm install
pnpm -r typecheckDevelopment discipline is defined in CLAUDE.md — hexagonal ports & adapters, DDD ubiquitous language, TDD, BDD. Contributions welcome: see CONTRIBUTING.md and the issue tracker.
Guild's code: Apache-2.0. Multica is a separate project under its own source-available license — review it before deploying.