MCP server
Connect Claude Code, Cursor and other AI tools
MindBacklog has a built-in MCP server (Model Context Protocol). Connect Claude Code, Cursor, claude.ai or
another MCP client and it can read your product: features and their Build Packs, stories, documents, problems,
feedback, the product map, competitors and what Mind knows. A few tools also write, and only as far as you could.
AI tools come with Growth and Scale (a trial of either included). The address is
https://app.mindbacklog.com/mcp.
Everything is in Settings › Your AI tools (the menu under your initials). The page has each tool's steps with the text to paste, and a first prompt to try: "Get the Build Pack for {KEY} from MindBacklog and build only what it lists. Ask me before going outside it."
Two ways to connect
Sign a tool in as you (no token to copy), for tools that support MCP sign-in:
- Claude Code: in your project folder,
claude mcp add --transport http mindbacklog https://app.mindbacklog.com/mcp, then type/mcp, pick mindbacklog and Authenticate. - Cursor: add the address (no token) to
~/.cursor/mcp.json; Cursor Settings › MCP shows Needs login: click it. - claude.ai: add it as a custom connector with the address.
Your browser opens MindBacklog: choose the workspace and one product (or every product you open), then Approve. The tool acts as you, with your access, checked on every call. It is listed under Signed in as you; Sign out refuses its next call. See Sign-in and security for two-step sign-in and how long a sign-in lasts.
Connect with a token, for tools that take one (scripts, most other clients). Make a token:
- What it is for, so you know which tool uses it ("Claude Code on my laptop").
- Which product: one product, or every product you open (the tool then names the product in each call).
- What it can do: Can read, or Can read and update (only where you can edit): it also moves a status, writes your product map and reports builds, as you could.
- Stops working after: 30 days, 90 days, 1 year or no end date; a Can read and update token lasts 30 or 90 days.
The token (it starts mb_) is shown once, right after you make it: MindBacklog keeps only a fingerprint. Keep it in an
environment variable or your tool's settings, never in a repository. You can have up to 10 active tokens in each
workspace at a time. The limit is per person: each teammate has their own 10. Your tokens lists each one with what it
can do, when it was last used and when it stops working; Revoke stops it on its next call.
- Claude Code with a token:
claude mcp add --transport http mindbacklog https://app.mindbacklog.com/mcp --header "Authorization: Bearer mb_…". With--scope projectthe token is written to.mcp.jsonin your repository: never commit that file. - Cursor: add the address and an
Authorization: Bearer mb_…header undermcpServersin~/.cursor/mcp.json(or use Add to Cursor), then ask for it in Agent mode. - Any other client that speaks MCP over HTTP: the address, and the token as an
Authorization: Bearerheader.
The tools
There are 35: 30 only read, 5 write. Name them exactly as below when you call one directly.
- The product:
list-products-tool,get-product-summary-tool,get-roadmap-status-tool. - Features:
query-features-tool,get-feature-details-tool,query-child-items-tool(a feature's stories and bugs),features-by-signal-evidence-tool(features ranked by the feedback behind them),search-work-items-tool. - Building:
get-build-pack-tool(what to build for a feature, and only that: its spec, who asked, the initiatives it serves with their goal, time budget, prediction and check date, and what it touches on the map; see Build Pack),get-stories-tool(its stories and acceptance checks),get-document-tool(its documents, live: a document's Copy for an agent ends with this call). - Product map:
get-product-map-tool(what your product does and how it's built; the map's age counts the commits on your default branch since the last sync). - Problems:
query-problems-tool(keys like P-12) andget-problem-tool. A problem's older Intelligence Hub key still works in both and inget-feature-details-tool. - Feedback (the tools call a piece a signal):
query-signals-tool,summarize-all-signals-tool,summarize-feature-feedback-tool,find-signals-for-objectives-tool,query-themes-tool,query-product-learnings-tool. - Objectives and teams:
query-objectives-tool,query-teams-tool. - What MindBacklog knows:
get-product-memory-tool(the team's notes, learnings, product DNA and summary),search-product-knowledge-tool,list-context-sources-tool,list-channel-sources-tool,query-comments-tool. - Competitors:
get-competitive-intelligence-summary-tool,get-battle-card-tool,query-competitor-events-tool.
The five that write, with a Can read and update token or a tool signed in as you, held by someone who can edit:
update-feature-status-tool: moves a feature to another status.update-story-status-tool: sets a story's or bug's status as your agent works. A story that follows Jira is refused: change it in Jira.accept-feature-proposal-tool: makes the call on a problem waiting for one (a P-n or its older key): accept builds it, dismiss and note close it, merge folds its feedback into a feature you name.map-product-tool: proposes a map of how your product is built; it shows the plan first and applies it only after you agree (undo within 30 days).report-build-tool: reports what was built (the stories, the files changed, what a user can now do, what isn't done; optionally the approach, decisions, deviations, tests, dependencies added, known gaps, follow-ups, what the Build Pack missed, and the effort).
Neither the map nor a build report changes a status. A write is recorded under your name, as if you had made it in MindBacklog, and shows in the item's Activity. A viewer's AI tools only read. When a workspace admin turns on Let AI tools update status (Settings › Security), an editor's signed-in tool may also tick off stories and mark features shipped on a product that editor can only read, and nothing else. Nothing can be deleted over MCP.
The server also has 6 resource templates and 5 prepared prompts (such as weekly-signals-briefing-prompt).
Safety
- A token is shown once and stored only as a fingerprint; revoking it, or signing a tool out, stops it on its next call.
- A token for one product can't see the others.
- Every write is checked against your role on every call. Deactivating someone, or turning their AI tools off in Settings › Team, stops their tools and tokens at once.
- Starter and the demo can't use AI tools.
If something doesn't work
- The tool shows MindBacklog as failed, or with no tools. For Cursor, check
mcp.jsonis valid JSON; check the token starts withmb_, isn't revoked and hasn't expired. For Claude Code,claude --debugshows the error. - "Your AI tools are turned off in this workspace." An admin turns them on in Settings › Team (it's free: no seat). When you ask for a token while yours are off, the workspace's admins are told.
- A plan error. AI tools need Growth or Scale.
- Empty results. A one-product token sees only that product; with an every-product token the tool names the
product in each call (
list-products-toollists them). - "The token is read-only". Make a Can read and update token, or sign the tool in as you.
Questions
- How is this different from Ask Mind? Ask Mind answers inside MindBacklog; MCP gives the same product to the AI tools you already work in.
- Can my AI client change anything? Only the five tools above, only as far as you could: a feature's or a story's status, the call on a problem, the product map and build reports. It can't comment, delete or link feedback.
- Does it cost AI credits? No: the tools make no AI calls. Accepting a problem can start a spec draft if your product drafts one on accept (Settings › Product), as accepting it in MindBacklog does.
- Can teammates share a token? Each person should make their own, or sign their tool in as themselves.