MCP Server — Overview

The Pathbound MCP server exposes the REST API as a set of tools for Model Context Protocol clients. The most common use is connecting Pathbound to Claude.ai so you can ask questions like “what did this contact do this week?” or “draft a follow-up email to the visitors who hit /pricing yesterday” — and have the model read your CRM, propose actions, and queue them for approval.

Server URL

https://mcp.pathbound.ai

When to use MCP vs REST

Use…	When
REST API	You’re writing your own backend, integration, or scheduled job.
MCP server	You want Claude.ai (or another MCP client) to act on Pathbound data on your behalf.

The two surfaces hit the same backend and respect the same tenant boundaries. MCP authenticates via Pathbound’s native OAuth login (no API key); the REST API authenticates via API keys.

What MCP exposes

47 tools across these domains:

• Contacts — list, get, create, update, timeline, conversations, email threads, notes CRUD.
• Companies — list, get, create, update, notes CRUD.
• Identities — list configured senders.
• Agents — list, get, update, trigger, list runs, get run.
• Actions — list, get, create, update, approve, reject, review pending.
• Events — list, get, types, stats.
• Segments — list, get members.
• Batch & Aggregate — multi-entity lookups and server-side group-by.
• Composite — review_pending_actions for the consolidated approval queue.

See the full Tools Reference.

Connecting Claude.ai

1. In Claude.ai, open Settings → Connected Apps → Add custom MCP.
2. Enter the URL https://mcp.pathbound.ai.
3. Claude opens the OAuth flow — log in with your Pathbound account.
4. Approve the connection. You’re now ready to ask Claude questions about your Pathbound data.

The OAuth grant is per-user and respects your Pathbound permissions: anything you can do in the Pathbound dashboard is what the MCP tools will let Claude do (with one exception — agent-proposed actions still require approval; the tool will draft, not auto-send).

Approval-by-default

Tools that send outbound communication or modify data are configured with destructiveHint: true in their MCP annotations and default require_approval: true in their parameters. In practice, a typical Claude session ends up with a queue of pending_approval actions — you (or another approver) can flip through them in the Pathbound dashboard or via review_pending_actions.

To execute without approval, set require_approval: false explicitly in the tool call. This is a deliberate per-tool opt-out, not a default.

Limits

• The MCP server is stateless and rate-limited per user. Heavy bursts (more than ~30 tool calls per minute) may be throttled.
• Tool responses are bounded — get_contact caps at 20 events / 10 actions; get_agent_run in full mode caps at 25 contacts. Use aggregate_data and batch_get_* instead of paginating.
• Long-running operations (LinkedIn login, agent triggers) return immediately with a job ID. Poll the corresponding tool to see progress.