All servers / Docs / Product Brain MCP server
Product Brain MCP server
Product Brain — MCP server for AI-assisted product knowledge management
Tested by ConnectATDScore 86 out of 100. Tested 2026-11-18 on package version . Package: @productbrain/mcp
Score breakdown
| Runs | 18 / 20 | |
|---|---|---|
| Tool descriptions | 20 / 20 | |
| Package safety | 12 / 20 | |
| Maintenance | 20 / 20 | |
| Adoption | 16 / 20 |
What we saw
- Starts and answers the MCP handshake: yes (185.7 seconds including download)
- Tools offered: 16
- Last published: unknown
- Weekly downloads (npm): 28,362
- License: none listed · Direct dependencies: unknown
Want this server set up for your team?
ConnectATD builds AI automation and wires up MCP servers like this one. Tell us what you are trying to do.
Talk to ConnectATDTools and how they are described
Agents choose tools from these descriptions. Under each name you can see how many of its parameters are explained.
session6 of 6 parameters describedStart here. When the person says "Start my Product Brain session.", call start. Session lifecycle, wrapup review, and work package facilitation. - **start**: Begin a tracked write session for attribution. **This is the ONLY call that creates a session** — write tools (capture, entries action=update/accept, relations) require it. Its reply is the briefing; then `orient task="<what you are doing>"` loads the context for a task. - **close**: End the current session. Records structured data (entries created, modified, relations). Run `session action=wrapup-review` before closing. - **status**: Check current session — active or not, oriented, activity counts. - **wrapup-review**: List this session's drafts with quality scores; accept each with entries action=accept. Run before close. - **resume**: Load session state (work package data, linked entries, drafts) from an existing work package entry.
entries23 of 23 parameters describedWrite entries on the Chain: the entry lifecycle after capture. Reads live on `context`. - **update**: Change fields on an existing entry (publishes whenever the entry is not draft — includes archived, not just current entries; autoPublish is accepted but ignored). - **accept**: Accept a draft entry onto the Chain. Runs the possible-clash check first. - **move**: Relocate an entry to a different collection; pass canonicalKey too for an atomic move+retype. - **verify**: Mark verified (lightweight, no scan — see `quality action=verify-chain`). - **replace**: one accepted entry replaces another of its kind, with a reason. - **still-true**: confirm accuracy (the holder of the owning role; if that role is empty, the nearest seated Team Lead up the team tree). Read entries with the read-only `context` tool: `context action=get entryId="..."`, `context action=search query="..."`, `context action=preflight`.
relations8 of 8 parameters describedCreate and delete relations between entries, each written only on request with exactly the verb you send. Read relations with the read-only `context action=find`. Each level links to the one above: A work package validates an accepted key result or KPI. A key result hangs on an accepted objective with part_of. An objective hangs on an accepted strategy with part_of. A strategy hangs on an accepted vision with part_of. A vision hangs on an accepted purpose with part_of.
context25 of 25 parameters describedRead the Chain. Every action is a read except incremental, which records the brief run it steers. - **get**: Full entries by ID: data, labels, relations, history. entryIds reads up to 20. - **batch**: get with entryIds. - **history**: Audit trail for an entry. - **list**: Browse entries by collection, status, tag or label; collections action=list names the slugs. - **search**: Full-text search; scope by collection or status. - **find**: An entry's direct relations, both ways; not text search. - **people**: The workspace's members: name, email, status, role. Max 200. - **outcome**: A key result's derived view (current, progress, pace, staleness, completion), its objective and up to 5 sibling key results. - **readings**: A key result's readings, newest first; a replaced row names its replacement. Record with `outcome action=record`. - **accountable**: What a role owns, reports on, and who holds it; no roleId: every seat. - **preflight**: What an entry in a collection needs; saves nothing. - **gather**: Three modes — (1) By entry: traverse the graph around a specific entry. (2) By task: load task-relevant governance, binding constraints, and supporting domain knowledge for a natural-language task. (3) Graph mode (entryId + mode='graph'): enhanced traversal with provenance paths. - **build**: Structured build spec for any entry — data, related entries, rules, glossary terms, chain refs. Use when starting a build. Requires active session. - **neighborhood**: Typed graph neighborhood — blocking chain (what this blocks, 3 hops), dependency chain (what blocks this), parent context (sibling count), related tensions, staleness signal, isolation signal. Fast indexed lookups. Requires entryId. - **changes**: Detect entries modified and relations created since a timestamp. Use for incremental updates instead of full traversal. Requires 'since' (ISO 8601). Returns grouped counts by collection. Cap: 200 entries. - **chain**: Directed deep traversal along one relation type to depth 4. Use for dependency analysis, governance chains, strategic alignment. Requires entryId. Optional: direction (outgoing/incoming), relationType, maxHops (1-4). Cycle detection, fan-out cap 10/hop, total cap 100 nodes. Returns tree + edges. - **cross-cut**: Structural aggregation — all relations of a given type across the workspace, grouped by source collection. Use to answer 'show me everything connected by part_of'. Requires 'relationType'. Cap: 500 entries. - **incremental**: Delta since last brief run for a skill. Returns only entries changed since the skill's last brief, filtering out previously surfaced entries. Requires 'skill' parameter. Use for steering briefs that need incremental awareness. - **brief**: Compound intelligence query composing multiple primitives. Requires 'briefType': 'steering' (7d changes + structure + delta + readiness), 'confidence' (7d changes + active work packages — tensions: use the pb CLI's `tensions list`), or 'delta' (changes since custom timestamp). Auto-records brief run for steering.
collections23 of 23 parameters describedManage knowledge collections and workspace labels. Eleven actions: - **list**: Browse all collections — glossary, decisions, rules, tensions, etc. Returns slug, name, description, and field schema. Use before capture to see what exists. - **describe**: Full documentation for a single collection. Returns field help text, option decision guides, usage guidance, examples, and cross-references. Use when you need to understand a collection's purpose, field semantics, or option values. - **update**: Update a collection's purpose, navGroup, fields, or (non-system only) stageLabelMap; its name, description, axis, status words and icon are the catalog's. Only provide the fields you want to change. Creating a NEW collection is not offered here — ask a platform admin to provision it, or enable the module ontology pack that owns the drawer. - **audit**: Health report for all workspace collections. Checks classification metadata, displayHint coverage per field, and field schema gaps vs system definitions. Returns total collections, count with issues, and per-collection issue list. - **export**: Full system_collection_definitions export as JSON. Includes all classification metadata (thinkingLayer, classificationPriority, classificationCheck, timelineRole, governanceRole, governanceFunction, etc.). Admin only. - **label-list/label-create/label-update/label-delete/label-apply/label-remove**: manage workspace labels (absorbs the `labels` tool) — list all labels, create/update/delete a label, or apply/remove a label on an entry. Labels can be applied to any entry across any collection for cross-domain filtering.
orient8 of 8 parameters describedTask-grounded context loader. Call `session action=start` first; it opens the session and hands you the briefing. Pass `task` to load the governance and context for a specific task. Orient is advice for captures, relations, accepts, ticks and questions; workflow checkpoints and organisation-structure writes (membership, ownership, parent) still require an oriented session. Purpose, vision and strategy highlights are always served in full; a grounding status flags when other context was reduced. Set depth with `tier` or `mode`; `rules` adds each rule's whole body.
capture16 of 16 parameters describedThe tool for creating a single knowledge entry. Creates an entry and returns a quality scorecard — all in one call. Provide a name and description; `collection` is optional — the kernel classifies when omitted and files a low-confidence pick as a draft with one line saying so. An agent capturing several atoms from one text calls this once per atom. **Explicit data:** When you know the schema, pass `data: { field: value }` to set fields directly. A value in `data` is kept as sent; `description` is the entry's text and fills the collection's prose field only when `data` leaves it empty (the same rule as `pb capture`); when `data` carries the prose field, `description` and `context` are used for classification and judging only and are not stored. Governance entries also require a top-level `why`. **Compound capture:** Pass `links` to create relations in the same call. An entry id typed in the text is linked as `references` and named in the reply; a capture writes no other relations. A capture is a `draft` until `entries action=accept`; questions use `question action=create`. The reply names possible duplicates; only an exact name in the collection is refused.
quality8 of 8 parameters describedAssess, re-evaluate, and verify entry/work package quality. Four actions: - **check**: Read the entry's server quality verdict (tier + criteria) — it does not re-score. - **re-evaluate**: Trigger a fresh quality evaluation. Returns heuristic verdict immediately, schedules LLM evaluation in background. - **verify-chain**: Verify knowledge entries against the actual codebase (codeMapping drift, cross-references) — absorbs `verify`. Produces a trust report with a trust score. - **spine-check**: Checks the direction, purpose to key results, for duplicates, unlinked entries and a strategy that refuses nothing; story:true also runs the one-story judgment (never scheduled, result labelled [unvalidated]). Async: a plain call reads back the latest stored verdict and auto-schedules a new run ONLY when none is already in flight AND the stored one is absent/stale/missing the requested story pass — repeated polling never re-triggers work. Pass refresh:true to force a fresh run.
workflows13 of 13 parameters describedChainwork workflow discovery, start/resume, checkpointing, and run inspection. Four actions: - **list**: Browse available workflows — retro, shape-a-bet, implementation-review, process-change. - **start**: Start or resume a workflow. Call with workflowId (e.g. 'implementation-review', 'retro', 'shape', 'process-change'). Returns the first (or current) round instructions and the exact checkpoint call to use when the round is complete. Use this when the user says 'run implementation review', 'run retro', 'shape this', or 'process change' — it is the tool entry point for running a workflow. - **checkpoint**: Record a round's output. Persists progress. At completion, some workflows support isFinal=true to create the summary chain entry. - **get-run**: Inspect a persisted workflow run by run ID or by workflow ID in the current session.
workspace4 of 4 parameters describedDiagnostics, session identity, and usage for the current workspace. Six actions: - **check**: Verify connectivity — workspace, collections, entries, latency. - **whoami**: Session identity — workspace ID, slug, name. - **status**: Workspace readiness — checks (pass/fail), gaps, stats (entries, relations, orphans). - **audit**: Session audit log — last N backend calls with summary. - **self-test**: Validate all registered tool schemas. - **usage**: LLM usage and cost summary — spend, tokens, model/feature breakdown.
feedback20 of 20 parameters describedSubmit Product Brain feedback from any agent, including with a read-only key and no session. list reads back your own workspace's feedback (all statuses, no gate). Vendor operators can also queue (triage), note, group, and change queue status.
conflicts3 of 8 parameters describedList possible clashes; propose; a person answers.
question12 of 12 parameters describedRouted questions — ask a role with `toRole` on create and the question is assigned to it at once. With no role named, your own team's Team Lead holds it; a Team Lead named without a team is your own team's. Six actions: - **create**: ask a question from text; `toRole` assigns it to that role at once; `toTeam`: The team whose Team Lead holds it; only for Team Lead, which is held in every team. - **assign**: reassign to a role (`toRole` required, `toTeam` as on create); the asker may, once, while a team lead holds it. - **answer**: writes `text` on the question as its answer, in full; `answerEntryId` may link an entry beside it, never instead. - **add**: Add a line to a question you asked; nothing already written changes. `entryId` and `text` required. - **force-close**: Team-Lead-only escape hatch (reason required) for a question that cannot otherwise be resolved. - **list**: what you owe (default); `mine:true` what you asked; `role` what waits on that role. `blocks` (repeatable, on create) records what a question is holding up — pass entry ids it blocks. While a question holds a draft, only a person accepts or closes. Its assigned role answers, unless that role asked or owns the draft.
role14 of 14 parameters describedWorkspace roles — who holds a role, where it sits and what it owns. Eight actions: - **seat**: put a person in this role, by person id or email, or by `newPersonName` for someone with no login yet (a placeholder person, created once per exact name). The role's home team is used; name `teamId` only for Team Lead. A held role refuses a second holder and names the holder; the person keeps their other roles in the team. Needs the team's Team Lead. With `replace`, the lead above replaces a team's Team Lead in one write. The lead directly above a team also decides who leads it. With `onMyAuthority`, the team's Team Lead seats a second Team Lead on their own authority; unseating them revokes it. - **unseat**: take this role from a person; their other roles in the team stay. The last active Team Lead of a team is refused. Invite a person, or take them out of a team, with the `people` tool. - **release_move** / **accept_move** / **refuse_move**: move a role or team (`childId`) to a team whose Team Lead is another seat. Acting as Team Lead of its current team, release it toward `teamId`; acting as Team Lead of that team, accept it, or refuse it with `why` (it stays, and a tension goes to its current team's Team Lead). - **move_work**: move every open item `fromRoleId` owns (drafts, current entries, open questions) to `toRoleId`, each checked as a one-entry ownership change; an item the receiver may not own goes to the nearest Team Lead of the old role's home team, and past items stay as history. Each call moves a capped number and says how many remain; call again to finish. `dryRun` lists the open items (name, kind, status, summary), 200 a page with how many of how many are listed, decides no destination and changes nothing; pass `page` for the next. - **give_right** / **withdraw_right**: Give or withdraw the organisation-design right on the team you lead. Read what a role owns with `context action=accountable`.
people6 of 6 parameters describedPeople in the workspace. Two actions: - **invite**: send a login link to `email`. Give `person` (a placeholder's id, as `role action=seat` printed it) to invite that person in place: they keep every role they hold. Or give `newPersonName` to invite someone new. A person who already has a login or an invite is refused. Needs a workspace owner or admin. - **remove**: take `person` out of the team `teamId`; every role they hold there goes with it. To take one role, use `role action=unseat`. Needs the team's Team Lead.
outcome13 of 13 parameters describedKey-result outcomes: record the measured readings behind a number. - **record**: Record a reading — evidence, not a verdict; any connector may record one. Read with the read-only `context` tool: `context action=outcome entryId="..."` (the derived view, its objective and sibling key results), `context action=readings entryId="..."` (the readings).
Tested in a sandbox with an empty environment: no API keys, a throwaway folder, nothing from a real account. A passing test shows the server runs and responds. It is not a security audit.