Expedify over MCP
Expedify speaks MCP (the Model Context Protocol) — so an AI agent (Claude Desktop, Claude Code, Cursor, n8n, or your own) can drive the whole product through one scoped API key. Search your CRM, author and run workflows, query analytics, ask your knowledge base — as tools the model calls directly.
MCP is a co-equal surface with the REST API, not an add-on. Both are projections of the same core:
┌──────────────┐
REST /v1/* ───▶│ │
│ ONE core: │ entity registry · scopes · services
MCP /v1/mcp ───▶│ operations │
│ .execute() │
└──────────────┘
Every guarantee you get on REST — the agx_ key, least-privilege scopes, default-deny, tenant isolation, rate limits — applies identically on MCP, because a tool call runs through the exact same gate as the REST endpoint.
Why MCP (and not just REST)
- Agent-native. Tools are curated intents (
search_contacts,run_workflow,ask_knowledge_base) that read like instructions — not a 1:1 mirror of every REST route. A model picks the right tool without you writing glue code. - One credential. The same
agx_key authorizes the whole MCP server. Its scopes decide which tools the agent is allowed to run (the catalog itself is public). - No SDK needed. MCP clients speak the protocol natively — connecting is a couple lines of config (see Connect).
The mental model
| REST | MCP | |
|---|---|---|
| Transport | HTTPS request/response | JSON-RPC (tools/list / tools/call) |
| Endpoint | https://api.expedify.ai/v1/... | https://api.expedify.ai/v1/mcp |
| Auth | Authorization: Bearer agx_… | Authorization: Bearer agx_… (same key) |
| Unit | an endpoint + method | a tool |
| Shape | fine-grained CRUD | curated agent intents |
| Discovery | this OpenAPI reference | tools/list (full catalog; scopes enforced at call) |
Start here
- Connect — point Claude Desktop / Claude Code / Inspector / n8n at the server (copy-paste).
- Auth & scopes — how a key's scopes gate which tools can run (discovery is public).
- Tool catalog — all 112 tools, grouped by module, with schemas + examples.
- Workflow tools — expose one of your own workflows as its own
run_…tool. - Recipes — worked end-to-end flows (author a workflow, ask a KB, NL→SQL→chart).
tools/list needs no auth (it returns the static catalog — no org data). Execution (tools/call) requires your key, and every call is scope-, tenant-, and ACL-checked.