Connect an AI tool through MCP, or export complete stored activity rows to your warehouse through the REST API.
Barry observes how your organisation works and builds a living process map — the company brain. Two ways in:
MCP — the fastest path. Point any MCP-capable client (Claude, ChatGPT, Cursor, VS Code) at your Barry server and query the brain conversationally.
REST API — for your own dashboards, pipelines, and integrations. Organisation-scoped exports in paginated JSON.
Your organisation gets a dedicated MCP server. Step-by-step guides for Claude, ChatGPT, Microsoft Copilot and other clients, plus the OAuth and tool reference, are in the MCP documentation.
Add it to Claude in one command:
Settings → Connectors → Add connector → paste https://getbarry.ai/mcp and authorise with your workspace SSO.
Create an organisation API key in the dashboard’s API Keys settings. Send it as a bearer token:
Organisation keys begin with sm_live_. The key determines the organisation; callers cannot override it. Empty Permissions and AvailableModules on the machine session do not prevent exports. MCP uses separate credentials or OAuth.
Use these routes, not an /api/v1/ prefix. See the full API documentation for pagination, late uploads, examples and the interactive OpenAPI reference.
| Method | Path | Returns |
|---|---|---|
| GET | /api/Export/activities | Complete EmployeeActivities rows and ActivityLog JSON |
| GET | /api/Export/employees | Employee rows, including stored device metadata |
| GET | /api/Export/organisation | The key’s organisation row |
| GET | /api/User/me | Validate the API key and inspect its machine session |
For activities, supply from=2026-10-08&to=2026-10-09 for one UTC day (end exclusive), or updatedSince without date bounds to catch late uploads across all activity dates.
Follow NextCursor until null, keeping all filters unchanged. Upsert whole rows by Id. Device metadata lives on employee rows; credentials are not exported.
Use your MCP client’s tool discovery and the MCP documentation for the available tools and authorisation requirements. For scheduled warehouse ingestion, use the dedicated export API.
ActivityLog is returned as stored JSONB, preserving every stored event kind and attribute, including custom fields such as ticket numbers. There is no event-kind filter or aggregate-only projection. JSONB does not retain original upload whitespace or object-key order.
The feed contains the organisation, employee and activity tables described above. It is a live table export, not a frozen snapshot or a change-data-capture log. See the incremental sync guide for checkpoint overlap and reconciliation.
Exports allow 2 concurrent requests per organisation and 4 globally. Page size is 1–100 stored rows (default 10), with an 8 MiB soft page budget; a single oversized row is returned whole. Database statements time out after 20 seconds.
MCP defaults to 60 requests per minute and 2 concurrent requests per credential and organisation; operators can override these defaults. For 429, honour Retry-After and retry with backoff. See limits and error handling.