Tell your agent what to build. Keelmarker keeps a checklist of everything it promised, ties each item to a test, and will not call the work finished until the tests pass. You see what holds up, what does not, and what is running in production. That is how you decide what to trust.
Copy this into Claude Code or any AI coding agent. It connects to Keelmarker, sets up your first product, and follows the gate from then on. Nothing to install.
Set up Keelmarker for this project, so that nothing counts as done until tests prove it. 1. Ask me for my Keelmarker access token (I create it under Settings in the app at https://keelmarker.com/app/). Then connect, replacing TOKEN with the token I give you: claude mcp add --transport http keelmarker https://keelmarker.com/api/v1/mcp --header "Authorization: Bearer TOKEN" If you are not Claude Code, add an MCP server with that URL and that Authorization header. Tell me if a restart is needed before the tools appear. 2. Call at_list_products. If this project has no product yet, call at_create_product with a short code and a name. 3. Call at_list_packs, tell me which starter packs fit this project, and apply the ones I choose with at_apply_pack. 4. For each requirement, write a test, then record it with at_create_test and link it with at_link_spec_test. 5. Run the tests and report the results with at_report_results, including the commit. 6. Before you tell me anything is done, call at_check_gate. If it is blocked, fix what it lists. Never say it is done while the gate is blocked.
Works with Claude Code and any tool that supports MCP over HTTP. You need an account to create the access token the agent asks for.
A simple loop that your agent follows and you can audit.
Before coding starts, the agent writes your request down in plain words.
Your request becomes specific statements that either hold or do not.
Each promise points at the tests that prove it. A promise that nothing tests shows as open.
The work cannot be closed until every promise has passing proof.
Agents are fast and confident. Keelmarker is how you know what to trust.
The verdict comes from test results tied to a specific commit, not from the agent’s summary.
Packs add the checks that are easy to forget: web security, payments, accounts and passwords, safe delivery.
Record which commit is running, so the verdict describes production and not just a branch.
Accepting a known risk is a human action with a record, never something an agent can do for itself.
Every request, promise, result and sign-off is kept, so you can see how a decision was made.
Command line, an MCP server and hooks let agents report as they work.
Sign-up is closed while we are small. Leave your email and we will write to you when there is room.
We use your email only to write to you about access. Privacy Policy