Integrations

Jelliu agents act on other systems through tools. A voice or chat agent can look up a customer in your CRM, book on your calendar or run an automation while the conversation is still going. Those tools come from three places:

Information also flows the other way, from Jelliu to your systems, through webhooks and Zapier triggers.

How it works

Tools are resolved when the agent runs, on voice and on text channels alike. A tool call acts immediately with the credential of the connection, so the permissions of the connected account are the outer limit of what an agent can do there.

Connected apps

Connect apps in the dashboard under Integrations. Each app is authorized with its own sign-in or API key, and the connection belongs to the workspace.

Once an app is connected, its tools are made available to every agent in the workspace. Connecting an app is the request for your agents to use it; there is no separate step per agent. To narrow that, use per-agent access.

Connection status

The Integrations screen shows each connection’s state:

Reconnecting an app that is already connected, to refresh an expired grant, does not count as a new app against your plan.

Writing interactions back to your CRM

When a connected CRM holds the contact, Jelliu writes each finished interaction (a call or a text conversation) back to it. The result is reported to your webhooks as crm_sync.completed or crm_sync.failed:

{
  "event": "crm_sync.completed",
  "timestamp": "2026-09-14T15:43:11.024Z",
  "data": {
    "channel": "voice",
    "source": { "kind": "call", "callId": "5b0d2f7e-9a41-4c3e-8f0a-2c6d1e7b9a10" },
    "callId": "5b0d2f7e-9a41-4c3e-8f0a-2c6d1e7b9a10",
    "conversationId": null,
    "agentId": "7c1d9f30-5e2a-4a7e-9c8f-91a0b1c2d3e4",
    "campaignId": "0f3c8b52-6d1a-4e2f-9b7c-4a5e6d7f8a90",
    "contactId": "a1b2c3d4-e5f6-4a7b-8c9d-0e1f2a3b4c5d",
    "results": [
      { "provider": "hubspot", "success": true, "externalId": "18273645", "error": null }
    ]
  }
}

A write that fails temporarily is retried before it is reported, so you do not receive crm_sync.failed followed by crm_sync.completed for the same interaction. A write that was already made by an earlier attempt counts as a success. To build your own writeback instead, see CRM sync with webhooks.

Importing contacts from a CRM

Filling a campaign’s call list from a connected CRM is done in the dashboard, from the campaign’s contact import dialog. To load contacts through the API, use the campaign contacts endpoints described in Contacts.

Control access per agent

Four fields on PATCH /api/agents/{agentId} decide what an agent may do with the workspace’s connected apps and MCP servers. Apps are named by their lowercase catalogue slug, such as hubspot or gmail.

curl -sS -X PATCH "https://api.jelliu.co/api/agents/7c1d9f30-5e2a-4a7e-9c8f-91a0b1c2d3e4" \
  -H "Authorization: Bearer $JELLIU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "allowedToolkits": ["hubspot", "gmail"],
    "toolkitAccess": { "gmail": "read" },
    "disabledTools": { "hubspot": ["HUBSPOT_DELETE_DEAL"] }
  }'

A write key is enough. Slugs are not checked against the catalogue: naming an app the workspace has not connected is harmless, because an agent only ever receives apps that are actually connected. The new tool surface is applied to the agent in the background after the update returns.

Custom MCP servers

Register an MCP server once per workspace with /api/mcp-servers, then assign it to the agents that should use it.

The MCP server object

The list endpoints (GET /api/mcp-servers and the per-agent list) return the same fields without toolOverrides. The secret token and custom headers are never returned by any endpoint.

Create a server

curl -sS -X POST "https://api.jelliu.co/api/mcp-servers" \
  -H "Authorization: Bearer $JELLIU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Inventory",
    "description": "Stock levels and delivery dates. Use it before promising availability.",
    "serverUrl": "https://mcp.example.com/mcp",
    "secretToken": "your-server-token",
    "transport": "STREAMABLE_HTTP",
    "approvalMode": "no_approval"
  }'

Response 201:

{
  "data": {
    "id": "e2b7c4d1-6f3a-4b8e-9c0d-1a2b3c4d5e6f",
    "tenant_id": "9f0fafb4-1c2d-4e5f-8a9b-0c1d2e3f4a5b",
    "agent_id": null,
    "elevenlabs_mcp_id": "mcp_8Qx2Lr7Tn4Vw1Zp6",
    "name": "Inventory",
    "description": "Stock levels and delivery dates. Use it before promising availability.",
    "server_url": "https://mcp.example.com/mcp",
    "approval_mode": "no_approval",
    "transport": "STREAMABLE_HTTP",
    "created_at": "2026-09-14T16:00:00.000Z",
    "updated_at": "2026-09-14T16:00:00.000Z",
    "toolOverrides": {},
    "provider": null
  }
}

Creating a server registers it with the voice engine before anything is saved. If registration fails, nothing is created and the API answers 502 with code MCP_SYNC_FAILED.

When you send toolOverrides and some of them could not be applied, the server is still created, the response status is 207, and a toolOverridesSync object (synced, removed, failed, failures) sits next to data. Only the decisions that were applied are stored.

Approval modes and per-tool decisions

There is no per-tool off switch. toolOverrides accepts auto_approved and requires_approval; sending disabled is refused with 400:

The voice engine has no per-tool off switch (only auto_approved / requires_approval). To stop an agent using create_invoice, detach this MCP server from the agent instead.

Per-tool decisions with any mode other than fine_grained are also refused, because they would have no effect:

Per-tool decisions only apply when the server's approval mode is 'fine_grained'; this one is 'always_ask'. Set approvalMode to fine_grained in the same request, or drop the per-tool decisions.

To stop an agent using a server’s tools, remove the server from the agent or delete the server.

Update and delete

PATCH /api/mcp-servers/{mcpServerId} accepts name, description, serverUrl, secretToken, customHeaders, approvalMode, toolOverrides and transport. Send at least one field, otherwise the response is 400 with At least one field must be provided. Changing serverUrl or transport re-registers the server with the voice engine; the id and agent assignments are kept.

DELETE /api/mcp-servers/{mcpServerId} first detaches the server from every agent, then deletes it, and answers 204. If one agent cannot be detached, the delete is refused with 502 MCP_SYNC_FAILED and the server stays in place, so no agent is left pointing at a server that no longer exists. Retry after a moment.

Assign servers to agents

A custom server does nothing until you assign it to an agent.

curl -sS -X POST "https://api.jelliu.co/api/agents/7c1d9f30-5e2a-4a7e-9c8f-91a0b1c2d3e4/mcp-servers" \
  -H "Authorization: Bearer $JELLIU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "mcpServerId": "e2b7c4d1-6f3a-4b8e-9c0d-1a2b3c4d5e6f" }'

Response 201. This endpoint answers with a message rather than a data object:

{ "message": "MCP server assigned to agent" }

The assignment is pushed to the voice engine in the same request. If that fails, the assignment is rolled back and the error is returned, so a 201 means the agent really has the server.

• GET /api/agents/{agentId}/mcp-servers lists the servers on an agent, in the list shape above. It can include Jelliu’s own tool gateway (mcp.jelliu.co), which the platform manages.
• DELETE /api/agents/{agentId}/mcp-servers/{mcpServerId} removes one server from the agent and answers 204.

Hosted automation providers

Zapier MCP, Make and Alegra’s MCP have fixed connection facts, so Jelliu offers them as presets. You supply the token; Jelliu sets the transport and approval mode, and attaches the server to every agent as soon as it is created.

GET /api/mcp-servers/providers returns the catalogue:

All three use STREAMABLE_HTTP; sending another transport is refused. A token is required:

Zapier needs the connection token from https://mcp.zapier.com

curl -sS -X POST "https://api.jelliu.co/api/mcp-servers" \
  -H "Authorization: Bearer $JELLIU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "provider": "zapier", "secretToken": "'"$ZAPIER_MCP_TOKEN"'" }'

Errors

See Errors for the envelope and validation details.

Limits

MCP servers per agent, not counting Jelliu’s own tool gateway:

Re-assigning a server that is already on the agent never counts against the cap.

Rate limits. The MCP server and assignment endpoints use the general per-workspace limit (see Rate limits). PATCH /api/agents/{agentId} shares the 10-per-minute budget for configuration changes.

Related