Integrations

MCP

Connect Codex, Cursor, Claude, and other coding assistants to Collabute context.

Use this guide to connect coding agents to Collabute via MCP.

Supported coding agents

Claude Code
Codex
Cursor
OpenCode
OpenClaw
Amp

Requirements

You must belong to an active Collabute organization.
Your role must be allowed to use the target integrations.
At least one integration should be connected in Settings > Integrations.
You need your MCP URL:
production: https://api.collabute.ai/api/mcp

Important auth note

Collabute MCP is OAuth-first. Do not configure static PAT headers for MCP clients.

When new scopes are introduced (for example meeting:read), you must re-run OAuth approval so your MCP client receives updated scopes.

Scopes

meeting:read: read recent meetings, meeting context, and transcript pages
memory:read: read memory retrieval data
linear:read: read cached Linear task data
slack:read: read indexed Slack thread/message data
vercel:read: read connected Vercel project/deployment data
write:propose: create pending proposals (no direct third-party write)

Capability catalog (for AI agents)

Use this section as the runtime capability contract for Claude, Codex, Cursor, OpenCode, OpenClaw, and Amp. At runtime, tools/list is the source of truth for what the current token can access.

| Tool | Scope | Use when | Returns | | --- | --- | --- | --- | | system.ping | none | You need a quick connectivity/tenant check | server status, org/user context | | meeting.list_recent | meeting:read | You need deterministic recency discovery for meetings | recent meeting list with transcript/summary flags | | meeting.get | meeting:read | You need actionable meeting context | meeting metadata + summary + memory evidence + cited transcript snippets | | meeting.get_transcript | meeting:read | User explicitly asks for full transcript content | paginated transcript segments (offset, limit) | | memory.search | memory:read | You need org memory retrieval by query | ranked memory items, traces, diagnostics | | memory.get_bundle | memory:read | You need intent-shaped context bundle | structured context + prompt addendum | | linear.search_tasks | linear:read | You need task lookup in connected Linear data | matching tasks list | | linear.get_task | linear:read | You need one exact task by id/key | single task payload | | slack.search_threads | slack:read | You need Slack thread/message evidence | matching messages + thread identifiers | | slack.list_suggestions | slack:read | You need generated Slack suggestions | suggestion list and statuses | | vercel.list_projects | vercel:read | You need project inventory | connected Vercel projects | | vercel.list_deployments | vercel:read | You need deployment status/history | deployment list | | vercel.get_deployment | vercel:read | You need one deployment detail | single deployment metadata | | meeting.propose_task_from_meeting | write:propose | User asks to create a task from meeting evidence | pending proposal only | | memory.propose_write | write:propose | User asks to persist memory updates | pending proposal only | | integration.propose_action | write:propose | User asks for integration-side mutation | pending proposal only |

Tool selection rules (for external agents)

Prefer deterministic retrieval tools before broad search.
For meeting-task requests, use this strict order:

meeting.list_recent -> meeting.get -> meeting.propose_task_from_meeting.

Do not call meeting.get_transcript unless user explicitly requests full transcript.
Use proposal tools for writes; do not assume direct mutation is available.
If a tool is missing in tools/list, the active OAuth token likely lacks that scope.

Prompt-to-tool playbooks

"Fetch my recent meetings and create a task out of it"

Call meeting.list_recent (pick top relevant meeting).
Call meeting.get with a focused query like "action items and owners".
Call meeting.propose_task_from_meeting with sourceSnippetIds from returned snippets.
Tell the user task creation is pending approval (proposal flow).

"Show me the transcript for meeting X"

Call meeting.get_transcript with offset: 0, limit: 200.
If hasMore = true, continue paging with nextOffset.

"What changed in Linear/Slack/Vercel recently?"

Use provider-specific read tool first:

linear.search_tasks, slack.search_threads, or vercel.list_deployments.

Use memory.search only when additional organizational context is needed.

Agent setup

Claude Code

Add the server:

bash

claude mcp add --transport http collabute https://api.collabute.ai/api/mcp

Start a Claude Code session and run:

text

/mcp

Complete the OAuth approval in browser.

Verify:

bash

claude mcp list

Codex

Add the server:

bash

codex mcp add collabute --url https://api.collabute.ai/api/mcp

Open Codex MCP management, then complete OAuth in browser when prompted.

Verify:

bash

codex mcp list

Cursor

Open Cursor MCP settings and add a remote MCP server:

name: collabute
url: https://api.collabute.ai/api/mcp

Or add URL-based config:

json

{
  "mcpServers": {
    "collabute": {
      "url": "https://api.collabute.ai/api/mcp"
    }
  }
}

Open a chat/tooling session and approve OAuth in browser.

OpenCode

Add a remote MCP server (interactive):

bash

opencode mcp add

Use:

name: collabute
type/transport: remote HTTP
url: https://api.collabute.ai/api/mcp

When OpenCode asks setup questions, use:

Location: Current project (or your preferred scope)
Enter MCP server name: collabute
Select MCP server type: Remote
Enter MCP server URL: https://api.collabute.ai/api/mcp
Does this server require OAuth authentication? Yes
Do you have a pre-registered client ID? No (recommended)

If your OpenCode build forces a client ID entry, use a stable value like:

text

opencode-collabute-local

Collabute MCP supports OAuth public clients (token_endpoint_auth_method=none).

Complete OAuth when prompted.

If needed, trigger auth manually:

bash

opencode mcp auth collabute

Verify:

bash

opencode mcp list

OpenClaw

Add Collabute MCP as a remote HTTP server:

bash

mcporter config add --transport http collabute https://api.collabute.ai/api/mcp

Run OAuth login for the server:

bash

mcporter config login collabute

Verify server and tool discovery:

bash

mcporter list collabute --schema

Do not set a manual Authorization header for Collabute MCP. OpenClaw/mcporter should use the OAuth flow and token cache.

Amp

Add the server URL in Amp MCP setup:

bash

amp mcp add collabute https://api.collabute.ai/api/mcp

Complete OAuth in browser when prompted.

Verify:

bash

amp mcp list

If your Amp build does not yet support remote MCP OAuth, update Amp first. Collabute MCP does not accept static PAT-style header setup.

Optional: manual device flow (advanced)

Use this only for clients/environments that cannot open browser auth automatically.

Start:

bash

curl -sX POST https://api.collabute.ai/api/mcp/oauth/device/start \
  -H "content-type: application/json" \
  -d '{"clientName":"Collabute MCP Client"}'

Open verification_uri_complete, sign in, and approve.

Exchange:

bash

curl -sX POST https://api.collabute.ai/api/mcp/oauth/token \
  -H "content-type: application/json" \
  -d '{"grant_type":"urn:ietf:params:oauth:grant-type:device_code","device_code":"<device_code>"}'

Refresh:

bash

curl -sX POST https://api.collabute.ai/api/mcp/oauth/token \
  -H "content-type: application/json" \
  -d '{"grant_type":"refresh_token","refresh_token":"<refresh_token>"}'

Verify installation

Confirm collabute is connected in your agent’s MCP list.
Ask the agent to list tools from collabute.
Run one read tool (for example meeting.list_recent) and confirm org-scoped data is returned.

Recommended first checks

meeting.list_recent
meeting.get
memory.search
linear.search_tasks
slack.search_threads
vercel.list_deployments

Meeting-first workflow

For prompts like "fetch my recent meetings and create a task out of it", agents should use this sequence:

meeting.list_recent to identify candidate meetings.
meeting.get to gather structured context plus cited snippets.
meeting.propose_task_from_meeting to create a pending Linear proposal.

meeting.get returns hybrid evidence by default:

ranked memory evidence
cited transcript snippets
no full transcript dump

If full transcript is explicitly required, use meeting.get_transcript (paginated with offset and limit).

When to use MCP

Use MCP when the request depends on organization context:

Team decisions, meeting memory, and follow-up context.
Live task/deployment/communication status.
Time-sensitive status questions.

For generic coding questions without org context, MCP can be skipped.

Common issues

Unauthorized or forbidden

OAuth approval was not completed.
Access token is invalid, expired, or revoked.
Your org role is restricted for one or more integration scopes.
Your token is missing meeting:read for meeting tools. Re-run OAuth consent to grant the new scope.

Tool returns no data

Integration is not connected.
Query scope is too narrow or incorrect.
No matching recent data exists.

Access denied in provider tools

Confirm your org integration policy allows your role.
Ask an admin/owner to update integration access settings.

Best practices

Start with read-first retrieval before action requests.
Use narrow prompts (team/project/channel/time window).
Keep one MCP connection per person/account.
Revoke access immediately when it is no longer needed.

Slack Troubleshooting