Collabute MCP

Use this skill when the task depends on private organizational context stored or indexed by Collabute.

When to use

Use this skill when requests include:

• Recent meetings, action items, owners, follow-ups, transcript content
• Team memory, prior decisions, historical context
• Linear task lookup from Collabute indexes
• Slack thread/message retrieval from Collabute indexes
• Vercel project/deployment status via Collabute integration
• “Create a task from this meeting” or similar write intent

Do not use this skill for generic coding questions that do not require workspace data.

Preconditions

1. The MCP server is already configured in the client as collabute.
2. Auth is OAuth-first; do not rely on static PAT headers for MCP.
3. Use tools/list as the source of truth for currently available tools/scopes.

Core behavior rules

1. Prefer deterministic retrieval tools before broad memory search.
2. For write intents, use proposal tools only; do not assume direct third-party mutation.
3. If a required tool is missing from tools/list, explain missing scope and ask user to re-auth.
4. If a meeting is referenced, keep all lookups scoped to that meeting before broadening.
5. Use transcript pagination for full transcript requests; avoid dumping large transcript output in one response.

Tool map

• system.ping
  • Use for connectivity and tenant sanity checks.
• meeting.list_recent (meeting:read)
  • Deterministic recent meeting discovery.
• meeting.get (meeting:read)
  • Structured meeting detail with summary, memory evidence, and cited transcript snippets.
• meeting.get_transcript (meeting:read)
  • Explicit full transcript access with pagination.
• memory.search (memory:read)
  • General memory-first retrieval across org context.
• memory.get_bundle (memory:read)
  • Intent-specific context packaging.
• linear.search_tasks / linear.get_task (linear:read)
  • Linear task lookup.
• slack.search_threads / slack.list_suggestions (slack:read)
  • Slack retrieval.
• vercel.list_projects / vercel.list_deployments / vercel.get_deployment (vercel:read)
  • Vercel retrieval.
• meeting.propose_task_from_meeting (write:propose)
  • Safe task creation proposal derived from meeting evidence.
• memory.propose_write, integration.propose_action (write:propose)
  • Proposal-only write flows.

Playbooks

Playbook A: “Fetch my recent meetings and create a task”

1. Call meeting.list_recent with a small limit.
2. Choose the best candidate meeting.
3. Call meeting.get with a focused query, such as "action items and owners".
4. Extract actionable items and corresponding snippet ids.
5. Call meeting.propose_task_from_meeting with:
   • meetingId
   • clear title
   • concise description
   • sourceSnippetIds when available
6. Tell the user the result is a pending proposal and requires approval.

Playbook B: “Show full transcript”

1. Call meeting.get_transcript with offset: 0 and bounded limit.
2. Continue while hasMore is true using nextOffset.
3. Summarize as needed and keep raw transcript output scoped to the user request.

Playbook C: “What changed in Linear/Slack/Vercel?”

1. Start with the provider-specific read tool (linear.*, slack.*, vercel.*).
2. Use memory.search only when additional cross-domain context is required.
3. Report results with timestamps/status fields when present.

Error handling

• 401 or unauthenticated:
  • Instruct user to complete OAuth auth flow for the Collabute MCP server.
• Forbidden scope (missing meeting:read, write:propose, etc.):
  • Ask user to re-authorize and grant required scopes.
• Not connected integration:
  • Tell user to connect integration in Collabute settings.
• No data found:
  • Suggest narrowing or broadening query by project/channel/time range.

Response contract for agents

When using Collabute MCP, responses should:

1. State which Collabute tool(s) were used.
2. Distinguish facts from inference.
3. Cite meeting snippet context when proposing meeting-derived tasks.
4. Clearly mark proposal outputs as pending approval.