MCP tools
These are the tools your agent sees after you connect your coding tool. The repo's AGENTS.md tells the agent to use them in this order.
The routine
- Start a session:
list_goals(what "shipped" means and the next milestone), thenstart_build_session. It picks the next ready task, moves its card to In Progress, and returns the build packet: acceptance criteria, files to touch, branch name and checks. - While working: after each commit,
record_build_progresswith a one-line summary. - Before committing:
validate_code_diffchecks the change against the accepted PRD and decisions. - Task done:
submit_build_evidence(checks run, commit) andcomplete_task. The card moves to Done. Thenstart_build_sessionagain. - Milestone done:
complete_milestone. Deployed:record_deployment, so the goal counts it. - Stuck:
report_blockerposts the question to Discord and your project page. Press Answer in Discord (or answer on the page); the agent receives it with its next Builderthread call, so you don't need to be at the computer it runs on. When it needs keys,request_env_valuestells you which file and how to get each one: add them, then press I've added them in Discord.
All tools
| Tool | Scope | What it does |
|---|---|---|
list_projects | projects:read | Lists the project this connection can use. |
get_agent_context | prd:read, build:read | Accepted PRD, decisions and open tasks. |
get_next_actions | build:read | Ready tasks in order, with what blocks the others. |
start_build_session | build:read, progress:write | Starts or resumes the next task and returns its build packet. |
get_build_packet | build:read | The packet for a specific task. |
record_build_progress | progress:write | Records progress on the current task. |
submit_build_evidence | progress:write | Attaches checks, test output or a commit to the task. |
complete_task | tasks:complete | Finishes the task and moves it to Done once a pushed commit is attached, the task's acceptance tests are written and CI is green on that commit. |
list_goals | projects:read | Goals, what counts as shipped, and milestone status. |
complete_milestone | tasks:complete | Marks a goal milestone done with evidence. |
report_blocker | progress:write | Asks the user for something the agent cannot do itself. |
request_env_values | progress:write | Asks you to put keys the agent cannot get itself into a local env file. You get what each is for and how to get it in Builderthread and Discord, and type the values into the file yourself. |
pause_task | progress:write | Puts a task back as not done while it waits on your answer, so the agent moves on. It comes back when you answer. |
get_user_answers | projects:read | Reads the user's answers to its questions. Answers also arrive appended to the agent's next call. |
record_deployment | progress:write | Records a working deployment (URL and commit). Never secrets. |
validate_code_diff | prd:read, build:read | Reviews a diff against the PRD and decisions before commit. |
propose_prd_change | changes:propose | Proposes a product change for you to approve. |
record_decision | changes:propose | Records a lasting technical decision. |
The connection's scopes are projects:read, prd:read, research:read, build:read, skills:read, changes:propose, progress:write and tasks:complete.
Plan limits
- Reading works for anyone on an active plan with MCP agent connections (Builder, Pro, and Studio).
- Starting tasks, recording progress, and calling
complete_taskorcomplete_milestonerequire Builder or above. - Every paid plan includes all build packets for your project with no monthly cap.
When a call is not part of the plan, the agent gets a clear "ask the user to upgrade" message and keeps coding locally.