MCP Server
Connic publishes an agent as an MCP tool with an auto-generated schema. Cursor, Claude Desktop, partner apps, and other agents can call it directly: same logic, no second implementation.
Overview
The Model Context Protocol is how AI applications discover and call external tools. The MCP Server connector puts Connic agents on the serving side of that protocol: every linked agent becomes a tool with a generated input schema, listed by a standard tools/list call and invoked by tools/call. Any MCP client, from Claude Desktop and Cursor to a partner's product or another team's agent, can use the agent without a custom API integration.
That makes one deployment serve three audiences at once. Internal teams call the agent from coding tools, partners call it as a capability inside their own AI products, and other agents compose with it at runtime. The agent's logic, tracing, and access control stay in one place; only the callers multiply. Compare the MCP server, connector, and client roles before deciding which integration direction fits.
How it works
Create the connector and pick a mode
Add an MCP Server connector from the agent's Connector Flow. Sync mode waits for the agent to finish and returns the result (up to 5 minutes). Inbound mode returns a run ID immediately and processes in the background, for jobs that take longer.
Link agents; each becomes a tool
Every linked agent is exposed as an MCP tool with a standardized schema: a required message and an optional structured payload the agent reads through template variables. Agent names map to tool names automatically.
Paste the endpoint into any MCP client
The connector generates a unique endpoint URL and secret key. Add both to the MCP client, and the linked agents appear in its tool list. Current and legacy MCP lifecycles are supported.
Production patterns
Patterns teams ship in production, with Connic running the queues, workers, and schedulers.
Partner calls the agent
mcp://your-product/agentEach partner gets its own credentials, rate limiting, and metering. Their AI tools call the agent like any other function. A paid API ships without a custom implementation.
Partner calls the agent
mcp://your-product/agentEach partner gets its own credentials, rate limiting, and metering. Their AI tools call the agent like any other function. A paid API ships without a custom implementation.
Main agent needs help
mcp://internal/pricingA pricing agent, an inventory agent, and a shipping agent each ship on their own. The orchestrator agent calls them as MCP tools, so each deploy stays independent and there is no monolith.
Main agent needs help
mcp://internal/pricingA pricing agent, an inventory agent, and a shipping agent each ship on their own. The orchestrator agent calls them as MCP tools, so each deploy stays independent and there is no monolith.
Customer connects their agent
mcp://customer/custom-logicCustomers register their own agent as an MCP tool inside the product. They extend the workflow with internal logic while the core experience stays consistent.
Customer connects their agent
mcp://customer/custom-logicCustomers register their own agent as an MCP tool inside the product. They extend the workflow with internal logic while the core experience stays consistent.
CRM agent needs billing data
mcp://billing/lookupThe CRM agent queries billing for payment history, and billing queries support for open tickets. Each team owns its agent, and MCP composes them at runtime.
CRM agent needs billing data
mcp://billing/lookupThe CRM agent queries billing for payment history, and billing queries support for open tickets. Each team owns its agent, and MCP composes them at runtime.
Authentication and access
Every request carries the connector's secret key, either as an Authorization: Bearer header or an X-Connic-Secret header. Authentication is on by default and can be disabled for trusted networks. Each connector has its own endpoint and secret, so a partner integration and an internal integration never share credentials, and revoking one touches nothing else.
Every tool call is a normal agent run underneath: it appears in run history with full traces, token and cost tracking, and the same guardrails and approval rules as any other trigger.
What a tool call looks like
On MCP 2026-07-28, the connector uses stateless requests and server/discover for capabilities, tools/list for the linked agents, tools/call to invoke one. The same endpoint preserves prior Streamable HTTP revisions with initialize and ping, plus the original GET/SSE transport for existing clients. A call sends a required message plus an optional payload object the agent accesses as template variables, so structured data arrives structured instead of packed into prose. See the full request and response formats in the MCP docs.
Information
- Publisher
- By Connic
- Connectors
- Connectors
- Modes
- Inbound, Sync
- Documentation
- MCP Server docs
Frequently Asked Questions
server/discover, tools/list, and tools/call; initialize and ping remain available only for legacy clients.Bring the event source, payload shape, result destination, and any private-network or approval requirements. We will map MCP Server to the right Connic connector mode, deployment path, and observability setup.
Talk to Sales