What we do
SpecStep generates the documentation your AI coder won't drift from. You answer a short interview — vision, users, requirements, constraints — and a team of specialist agents collaborates to produce a complete documentation package: vision, requirements, architecture, ADRs, phase plan, traceability matrix, session protocol. Hand it to Claude Code, Cursor, or Copilot and they ship without re-deriving context every chat.
Otto runs the interview. When your project hits the edge of his expertise — payments, accessibility, data modelling, security — he pulls in the right specialist live, then folds their input back into the spec.
Why this exists
AI coders are excellent at the next ten lines of code and unreliable at the next ten thousand. They drift. They forget. They invent things that already exist. The fix isn't a smarter prompt — it's giving the agent a written contract about what the system is, how it's organized, and what "done" looks like. Hand-writing that contract is slow and error-prone. SpecStep automates it without sacrificing the rigor.
How it works
The whole flow runs in roughly the time it takes to make a cup of coffee for short projects, or up to half an hour for the most thorough review profile. You can run it in the web app, through the API, or directly from Claude Desktop / Cursor / Copilot via the MCP server.
- Interview — Otto runs a focused conversation covering vision, users, scope, requirements, architecture, and constraints. Specialists join when their expertise is needed. You can pause, resume, and edit answers anytime.
- Generation — A team of agents drafts each file, self-reviews against the rubric, and (depending on the review profile) gets a fresh-eyes pass. Cost and duration estimates are shown up-front, grounded in your actual usage history.
- Delivery — The package downloads as a zip, commits to a GitHub repository, or returns over REST / MCP. Your AI coder reads it, follows it, and stops drifting.
Four review profiles
- Fast — one self-review round. Good for prototypes and proofs of concept.
- Normal — multiple self-review rounds plus a fresh-eyes pass. The default for production work.
- Extensive — the most thorough review with multiple fresh-eyes passes. For systems where the cost of getting it wrong is high.
- Researcher — runs Fast, Normal, and Extensive in parallel and lines the results up side-by-side. One Researcher credit, three lenses on the same intake. Useful when you want to see how depth of review changes the package.
Each profile shows a per-intake cost and duration estimate as you go — no hardcoded approximations, no surprise bills.
Three surfaces, one engine
The same generation engine ships behind three surfaces, all on one account, all sharing one rate limit and audit trail:
- Web — the interactive workspace at specstep.com.
- REST + Webhooks — a JSON API for CI integrations and scripts; signed webhooks notify your systems when a generation completes.
- MCP — a JSON-RPC server your AI coder calls directly. Claude Desktop, Cursor, Codex, GitHub Copilot, and other MCP clients sign in through your browser — no API key to copy and paste.
External connectors
Connect a SharePoint site, a OneDrive folder, or a Google Drive folder during the interview and SpecStep summarizes the contents and feeds them into your spec as reference documents — no manual copy-paste. Dropbox is next. Free accounts can connect and preview; generating from connector-sourced references is a paid feature.
Where we're going
The platform is in active build-out. Near-term: the Team plan, more specialist agents, and broader source-control coverage. We post user-facing changes to the release notes as they ship; drop us a line if you'd like to be notified about something specific.