Skip to content

AI assistants (MCP)

Baseplate runs an MCP server, so an AI assistant can read your sprint, create and move tasks, and write docs. It always acts as you, with your permissions, and every change it makes is logged as “via MCP”.

  1. In Claude, add a custom connector with the address https://{your Baseplate domain}/mcp.
  2. Claude opens Baseplate in your browser. Sign in, choose what it may do and which studios it can see, and select Allow.
  3. To disconnect it later, open Settings → API & apps.

Clients that support remote MCP servers and OAuth sign-in can use this config:

{
"mcpServers": {
"baseplate": { "type": "http", "url": "https://{your Baseplate domain}/mcp" }
}
}

Clients without OAuth can send a personal access token instead (Authorization: Bearer bp_pat_…). Create one in Settings → API & apps.

Tool What it does Permission
whoami Who you are and which studios you’re in read
list_projects, get_project Projects, workflow statuses, sprints and members read
list_tasks, get_task Tasks, filtered by status, sprint or assignee read
create_task, update_task, move_task, comment Change tasks tasks:write
get_folder_structure, list_assets, get_asset The asset library read
request_upload, finish_upload, submit_task Upload files and submit work for review submit
search Tasks, projects, docs and assets read
get_doc, create_doc, update_doc Docs as Markdown docs:write
list_my_invoices, draft_invoice Your invoices; drafts only invoices:draft
sprint_summary Points done, tasks by status, blocked work and each person’s load read
propose_plan, apply_plan Set up a project from a design doc: propose a plan to review, then create it tasks:write
bulk_create_tasks, link_tasks, create_sprint, assign_tasks Many tasks, Task Web links, a sprint or assignments in one call tasks:write
undo_batch Take back everything one of the calls above made tasks:write

Baseplate doesn’t run an AI for this, so it costs nothing: you use your own through MCP.

  1. Give your AI the game design doc and ask it to plan the project. If the doc is in Baseplate, open it and choose ⋯ → Plan with your AI… for a ready prompt that points your AI at the page (get_doc) and the right project. It calls propose_plan with epics, tasks (acceptance criteria, discipline, estimate, target folder, what each depends on, sprint and assignees) and sprints, which point at each other by short refs it picks. Nothing is created yet.
  2. Open the review link it gets back (Project settings → AI plans). Untick anything you don’t want, fix titles or estimates, and press Create. Planners only.
  3. Everything lands in one batch, and the tasks show From an AI plan. If it isn’t right, Undo on the AI plans page (or undo_batch) takes it all back, as long as nobody has started work on it.

Every write tool takes an optional idempotencyKey (or the Idempotency-Key header on the REST API). Send the same key when retrying, and you get the first answer back instead of a second copy of the work.

Docs are also listed as MCP resources (baseplate://{studio}/docs/{id}).

For a sprint update, ask your AI to summarise the sprint: sprint_summary gives it points done, tasks by status, what’s blocked and each person’s load.

  • There are no tools for payout details, and none that send an invoice or mark one paid. Those always happen in the web app.
  • update_doc replaces a whole page, so it needs confirm: true.
  • Each token can make 120 calls a minute.