The MCP surface here is deliberately tiny — a stateless reader over a fixed set of artifacts that returns the next command — because the actual work is done by a CLI and by the files it leaves in the repository. That is the right shape: the instruction files, hooks and ADRs keep working in every session and every assistant long after the server is gone. The score is the other half of the argument: fourteen points across seven properties with the evidence cited per line is a claim you can check, which is more than a linter score or a vibe.
A setup-time quality framework for AI-assisted projects. One command scans the repository, detects the stack and writes tailored instruction files for whichever assistants you use — Claude, Cursor, Copilot, Windsurf, Cline, Aider — alongside a status tracker, pre-commit hooks and spec skeletons. The MCP server itself is a small sentinel: it reads the generated artifacts, works out the correct next command, and returns it.
- Instruction files assembled from 116 curated blocks matched to your stack — SOLID rules, hexagonal architecture with ports and adapters, a testing pyramid with coverage targets, clean-code conventions, CI/CD stages, 12-factor operational rules and domain patterns — rather than one generic file.
- 24 domain-specific rule sets, so a fintech or healthcare project starts with the constraints that domain actually has.
- A score out of 14 across seven properties — self-describing, bounded, verifiable, defended, auditable and the rest — each line citing the evidence behind it, so a failing property names the file or the missing artifact.
- Quality gates organised by release phase, from development through pre-release hardening, release candidate, deployment and post-deployment, each with a condition and an evidence requirement — and gates that cannot be auto-passed because they require a human.
- Architecture decision records auto-sequenced in MADR format under docs/adrs, with context, decision, alternatives and consequences, so past choices stop being re-litigated.
- A readiness cascade the assistant can check: functional spec, architecture, constitution, ADRs and use cases, with the audit reporting named violations — file length, layer violations, hardcoded URLs and credentials, mocks in production code, missing specs, a stale status file — each mapped to a remediation workflow.
- Session continuity through a status file and a config file that persist context between sessions, and a refresh that detects when the project's scope has drifted from the spec.
- The sentinel tool itself: it reads the config, the instruction file and the hooks, derives the next command and returns it — nothing more.
Node, and a project directory to point it at — the setup command runs through npx with nothing installed. It works best with an AI assistant in the loop: without one it falls back to directory heuristics and says so. The important expectation is that this is a setup-time tool with no runtime footprint, and the sentinel says the same about itself: register it, get the project configured, then remove it to reclaim the context it occupies.
One command — npx -y forgecraft-mcp