Your coding agent writes the spec. This server makes it complete.
An MCP server and a CLI for Claude Code, Cursor, Windsurf and any MCP client. The agent reads the
questionnaire, answers it from your repository and the conversation, and writes AGENTS.md
and CLAUDE.md — schema, auth, payments, SEO, security, legal, file structure.
Claude Code
Remote server, nothing to install:
claude mcp add --transport http specpack https://prompt-generator-website.com/mcp
Or the npm package, which adds two offline tools (audit and init) that read your files locally:
claude mcp add specpack -- npx -y specpack mcp
Then ask: “Use specpack to write the AGENTS.md for this project.”
Cursor, Windsurf and other clients
Add this to .cursor/mcp.json (or your client's MCP config):
{
"mcpServers": {
"specpack": {
"url": "https://prompt-generator-website.com/mcp"
}
}
}
With an API key (free account, or a paid plan for unlimited specs), add a header:
"headers": { "Authorization": "Bearer pgw_…" }
The tools
| Tool | What it does | Runs |
|---|---|---|
get_questionnaire | The questions for a project type, with options and conditions. The agent answers them from what it knows. | server, free |
generate_spec | Answers in, full spec and agent files out. Deterministic: same answers, same bytes. | server, 1 spec |
draft_answers | One-line idea in, a complete answer set and the assumptions made. Uses Claude on our side. | server, 1 AI draft |
list_project_types, get_usage | Types, plan and quota. | server, free |
audit_agent_file | Bloat score of a CLAUDE.md / AGENTS.md / rules file: lines to cut, lines to move out of context. | your machine (npm) |
draft_agent_file | A lean AGENTS.md from what the repo already declares: scripts, package manager, framework, test runner. | your machine (npm) |
The CLI
npx specpack audit # CLAUDE.md, AGENTS.md, .cursorrules… in this folder
npx specpack audit --max-bloat 40 # exit 1 above 40 % bloat: put it in CI
npx specpack init # write a lean AGENTS.md from the repo (offline)
npx specpack spec --describe "A booking app for dog groomers with deposits"
npx specpack login pgw_… # use your plan instead of the free quota
audit and init never leave your machine. spec calls this site.
Source: github.com/THE-KIPDEV/specpack (MIT).
The API
Plain JSON, CORS open, no key needed for the free quota. Full contract: API-V1.md.
curl https://prompt-generator-website.com/api/v1/questions?type=saas
curl -X POST https://prompt-generator-website.com/api/v1/specs \
-H "Content-Type: application/json" \
-d '{"type":"saas","answers":{"project_name":"Ledgerly","project_description":"Invoicing for freelancers who bill in several currencies"}}'
Limits
Without a key: 2 specs and 3 AI drafts per address. 7-day Pass ($5 once): unlimited specs, 30 AI drafts. Pro ($9/month): unlimited specs, 100 AI drafts a month. Keys are created in your account.
Questions
Do I need an API key?
No. Without a key the server counts 2 specs and 3 AI drafts per address. A key (free account) ties usage to your account; a Pass or Pro plan makes specs unlimited.
Which MCP tools are there?
list_project_types, get_questionnaire, generate_spec, draft_answers and get_usage on the remote server. The npm package adds two local, offline tools: audit_agent_file and draft_agent_file.
Does the agent send my code to your server?
No code is sent. generate_spec receives the questionnaire answers the agent chose to send. audit and init in the npm package run entirely on your machine.
Why would my agent use the questionnaire instead of writing a spec itself?
Because the questionnaire is the checklist: payments, auth edge cases, GDPR pages, rate limits, file structure. The engine is deterministic, so the same answers give the same spec, and the spec is diffable.
Is the npm package open source?
Yes, MIT: github.com/THE-KIPDEV/specpack. It works offline for audit and init; spec and the remote tools call this site.