specpack MCP server

specpack

Lint and generate the files your AI coding agent reads — CLAUDE.md, AGENTS.md, .cursorrules, Cursor rules — and turn a project description into a full build spec. CLI, library and MCP server.

npx specpack audit          # bloat score for the agent files in this repo (offline)
npx specpack init           # write a lean AGENTS.md from your manifests (offline)
npx specpack spec --describe "Booking app for dog groomers with deposits and SMS reminders"

CLAUDE.md / AGENTS.md are loaded into the agent's context on every session. Every line that restates package.json, pastes a file tree or tells the model to "write clean code" costs tokens and competes with the two rules that actually matter. specpack audit finds those lines; specpack init starts you at the trimmed end.

Node ≥ 18. No account needed. audit and init never touch the network.

---

specpack audit — how much of your memory file does nothing?

Each line is classified keep, cut (removing it changes nothing: role-play preamble, generic advice, restated manifests, duplicates, emphasis inflation, prose) or offload (true but situational: file trees, schema dumps, endpoint lists, long code blocks, setup/deploy sections — move them to their own file and reference them with @docs/…). Same rules as the browser tool at prompt-generator-website.com/claude-md-audit.

Real output on a typical file:

$ npx specpack audit CLAUDE.md
CLAUDE.md  59% bloat  · 44 lines · ~341 tokens every session
  10 keep · 12 cut · 15 offload of 37 non-empty lines · ceiling 200, trimmed files land near 60
  Over half of this file is doing nothing. It is not neutral weight — it is what your real rules are competing against.

     3  cut     You are an expert senior full-stack developer with 10 years of experience.
        ↳ role-play preamble: CLAUDE.md is project memory, not a system prompt. "You are an expert…" costs tokens and changes nothing about the code produced.
     4  cut     Please always write clean, readable and maintainable code.
        ↳ default behaviour: Removing this line does not make the model write worse code — it already tries to do this. It only dilutes the lines that matter.
     5  cut     IMPORTANT: follow best practices at all times.
     …
    12  cut     - Next.js 15.1
        ↳ restates a manifest: package.json / composer.json / requirements.txt already state this, and the agent reads them. Keep it only if the version pin has a non-obvious reason.
     …
    18  offload ## Database schema
        ↳ situational section: Only a fraction of your sessions need this. Move the section to its own file and point at it with @docs/… so it loads when the task asks for it.
     …
    31  offload src/
        ↳ file tree: Root of a pasted directory tree. The agent can run ls and get the real one — a copy goes stale after the first refactor.
     …
  Trimmed copy (8 lines): specpack audit CLAUDE.md --trimmed

Without arguments it audits every agent file in the current directory: CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules, .github/copilot-instructions.md, CLAUDE.local.md, .cursor/rules/*.mdc. - reads stdin.

Option
--trimmed	print only the keep-only version (> CLAUDE.trimmed.md)
--json	every line with verdict, rule and reason, plus stats and the trimmed file
--all	also list the kept lines and why
--max-bloat <pct>	exit code 1 when a file is above the threshold

In CI

# .github/workflows/agent-files.yml
name: Agent files
on: [pull_request]
jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx -y specpack audit --max-bloat 30

specpack init — a lean AGENTS.md from what the repo already says

Reads package.json (scripts, lockfile → package manager, framework), tsconfig.json, the Prisma schema, pyproject.toml / requirements*.txt (uv, Poetry), composer.json, go.mod, Cargo.toml, Gemfile, Makefile, Docker Compose, .nvmrc, .env.example, .claude/commands. Only what a file states is written: nothing is guessed, no patch versions, no "You are an expert" preamble. Sections with nothing detected are left out.

Real output on a Next.js + Prisma repository:

$ npx specpack init
✓ Wrote AGENTS.md (37 lines, 0% bloat)
  Scanned from package.json, tsconfig.json, prisma/schema.prisma, .nvmrc, docker-compose.yml, .env.example, .gitignore, .claude/commands/.

# AGENTS.md — ledgerly

Invoicing app for freelancers. A web application.

## Stack
- Next.js (App Router)
- Tailwind CSS
- Zod
- Stripe
- Vitest
- Playwright
- ESLint
- TypeScript (strict)
- Prisma + PostgreSQL (schema: prisma/schema.prisma)
- Local services via Docker Compose (docker-compose.yml): db

## Commands
- `pnpm dev` — `next dev --turbopack -p 3001` (port 3001)
- `pnpm build` — `next build`
- `pnpm start` — `next start`
- `pnpm test` — `vitest run`
- `pnpm test:e2e` — `playwright test`
- `pnpm lint` — `next lint`
- `pnpm db:migrate` — `prisma migrate dev`
- `pnpm db:seed` — `tsx prisma/seed.ts`
- `docker compose up -d`

## Project rules
These are things the agent cannot guess from the code. Follow them exactly.
- Use pnpm (pnpm-lock.yaml), not npm or yarn — another package manager ignores the lockfile.
- Runtime: Node 20 (.nvmrc).
- Environment variables are listed in `.env.example`; the real `.env` is git-ignored — do not commit it.
<!-- Add the rules the code cannot tell the agent, one per line: the directory scheduled for deletion,
     money stored in cents, the command that must never run against staging. -->

Then add the rules only you know under Project rules — that is the part that matters.

Option
--format agents	AGENTS.md (default)
--format claude	CLAUDE.md — imports @AGENTS.md when it exists, plus Claude Code specifics (slash commands, .mcp.json servers)
--format cursor	.cursor/rules/project.mdc (always-applied rule)
--format all	the three
--dry-run / --force	preview / overwrite (existing files are never overwritten otherwise)

specpack spec — a full build spec for a new project

For a project that does not exist yet, prompt-generator-website.com turns a questionnaire (7 project types: SaaS, e-commerce, business site, marketplace, blog, web app, landing page) into a complete specification — stack, database schema, authentication, type-specific features, pages, design, SEO, security, legal, deployment, file structure — plus AGENTS.md, CLAUDE.md, .cursor/rules/project.mdc, .cursorrules, .windsurfrules and .github/copilot-instructions.md that keep the agent on-spec for the whole build.

# AI drafts the answers from a description (10–40 s), then the spec is generated
npx specpack spec --describe "Booking app for dog groomers with deposits and SMS reminders" --out ./groombook

# your own answers (see GET /questions in the API)
npx specpack spec --answers answers.json --type saas

# or answer the questionnaire in the terminal
npx specpack spec

The spec engine is deterministic: the same answers always give the same bytes. Free without an account: 2 specs and 3 AI drafts. More: a 7-day Pass or Pro at prompt-generator-website.com/pricing; then specpack login pgw_… (or SPECPACK_API_KEY). specpack whoami shows what is left.

MCP server

Seven tools for Claude Code, Cursor, Windsurf, VS Code and any MCP client:

Tool	Where it runs
audit_agent_file	local	bloat score, lines to cut/offload with reasons, trimmed copy
draft_agent_file	local	lean AGENTS.md / CLAUDE.md / Cursor rule from the manifests (returns it, writes nothing)
list_project_types	API	the 7 project types
get_questionnaire	API	every question for a type; the agent fills it from the conversation and the repo
generate_spec	API	spec + agent files, written to output_dir if given (files picks which). No AI on the server
draft_answers	API	AI fills the questionnaire from a description (uses an AI draft)
get_usage	API	plan and what is left

Claude Code

claude mcp add specpack -- npx -y specpack mcp

Cursor (.cursor/mcp.json), Windsurf, Claude Desktop

{
  "mcpServers": {
    "specpack": { "command": "npx", "args": ["-y", "specpack", "mcp"], "env": { "SPECPACK_API_KEY": "" } }
  }
}

Remote, nothing to install (streamable HTTP; audit_agent_file and draft_agent_file read local files, so they only exist in the local server): https://prompt-generator-website.com/mcp — optional header Authorization: Bearer pgw_….

claude mcp add --transport http specpack https://prompt-generator-website.com/mcp

Library

import { audit, scanProject, renderAgentsMd, SpecpackClient } from "specpack";

const { stats, lines, trimmed } = audit(fs.readFileSync("CLAUDE.md", "utf8"));
console.log(`${stats.bloat}% bloat`, lines.filter((l) => l.verdict === "cut").length, "lines to cut");

const agentsMd = renderAgentsMd(scanProject("."));

const client = new SpecpackClient({ apiKey: process.env.SPECPACK_API_KEY });
const draft = await client.draft("A booking app for dog groomers");
const spec = await client.createSpec(draft.type, draft.answers); // spec.spec, spec.files["AGENTS.md"]…

createMcpServer() is exported from specpack/mcp. API errors are SpecpackError with code (quota_exceeded, unauthorized, invalid_request, …), status and upgradeUrl.

Reference

• Exit codes: 0 ok · 1 failure, or --max-bloat exceeded · 2 usage error or quota exceeded.
• Environment: SPECPACK_API_KEY, SPECPACK_API_URL (testing), NO_COLOR.
• Key storage: ~/.config/specpack/config.json ($XDG_CONFIG_HOME honoured), mode 600.
• HTTP API: API-V1.md.
• Privacy: audit and init read local files and send nothing. spec sends your answers or description to prompt-generator-website.com.

License

MIT © Kipdev