MCP Server
Since Camel 4.22
The camel-mcp-server module exposes Camel routes registered via the ai-tool and ai-resource components as tools and resources of a Model Context Protocol (MCP) server, served over MCP streamable HTTP. No route is needed for the server itself: add the dependency, configure which tags to expose, and every matching ai-tool route becomes an MCP tool and every matching ai-resource route an MCP resource, discoverable by any MCP client (another Camel application, an IDE, a coding agent).
Maven users will need to add the following dependency to their pom.xml:
<dependency>
<groupId>org.apache.camel</groupId>
<artifactId>camel-mcp-server</artifactId>
<version>x.x.x</version>
<!-- use the same version as your Camel core version -->
</dependency> Architecture
The module is split in two artifacts:
-
camel-mcp-server-api— the runtime-agnostic bridge and the smallMcpServerEngineSPI. The bridge owns selection (tags), execution via the sharedAiToolExecutorandAiResourceExecutor(per-call timeout, error sanitization) and reacts toAiToolRegistryandAiResourceRegistrychanges when routes start and stop. It has no dependency on the MCP Java SDK. -
camel-mcp-server— the serving engine for Camel Main and Camel JBang, built on the official MCP Java SDK with a Vert.x streamable HTTP transport. The MCP endpoint is registered on the Camel main HTTP server’s router, so it serves on the main server port (camel.server.port) and inherits its lifecycle, authentication and CORS configuration.
Engine resolution mirrors the platform-http engine: a bean of type McpServerEngine in the Camel registry wins; otherwise the engine is discovered on the classpath. Other runtimes plug native engines through the same SPI: on Quarkus the camel-quarkus-mcp-server extension serves through the Quarkiverse quarkus-mcp-server (configured via quarkus.mcp.server.), and on Spring Boot the starter serves through the Spring AI MCP server (configured via spring.ai.mcp.server.). Bridge behavior — tag selection, timeout, sanitization — is identical on every runtime and verified by a shared conformance test kit.
Resource support is opt-in on the SPI: resourceAdded and resourceRemoved default to no-ops and an engine declares supportsResources() when it serves them. An engine that serves tools only keeps working unchanged, and the bridge logs a WARN when ai-resource routes match the tags but the engine cannot serve them. The conformance kit follows the same split: resource assertions live in a separate McpServerResourceConformanceTestSupport class that an engine extends once it supports resources.
Usage
Define tools as regular ai-tool routes and give them tags:
-
Java
-
XML
-
YAML
from("ai-tool:query_db?tags=crm" +
"&description=Query customer database" +
"¶meter.customerId=string" +
"¶meter.customerId.description=The customer id" +
"¶meter.customerId.required=true")
.to("jdbc:dataSource"); <route>
<from uri="ai-tool:query_db?tags=crm&description=Query customer database&parameter.customerId=string&parameter.customerId.description=The customer id&parameter.customerId.required=true"/>
<to uri="jdbc:dataSource"/>
</route> - route:
from:
uri: ai-tool:query_db
parameters:
tags: crm
description: "Query customer database"
parameter.customerId: string
parameter.customerId.description: "The customer id"
parameter.customerId.required: "true"
steps:
- to:
uri: jdbc:dataSource Resources
Read-only content is exposed as MCP resources through ai-resource routes, selected by the same tags:
from("ai-resource:app_config"
+ "?resourceUri=camel:///config/app.json"
+ "&tags=crm"
+ "&description=Current application configuration"
+ "&mimeType=application/json")
.pollEnrich("file:config?fileName=app.json&noop=true&idempotent=false", 5000); An MCP client then sees the resource in resources/list and fetches it with resources/read. A read carries no arguments: the route is invoked with an empty exchange and its body is the content. The mimeType decides the wire format — textual types are sent as text, everything else as a base64 blob — see ai-resource for the details.
Unlike tools/call, resources/read has no in-band error flag, so a failed read is returned as a JSON-RPC error with the same sanitized message the client would get from a failing tool.
Tool annotation hints
ai-tool routes can declare optional MCP ToolAnnotations hints (title, readOnlyHint, destructiveHint, idempotentHint, openWorldHint). The bridge passes them through to the MCP engine. See MCP Tool Annotation Hints for configuration examples. Hints are advisory UX metadata only — not authorization.
On Camel Main and Camel JBang no code is needed — like Jolokia or Prometheus, the server starts from configuration properties alone:
camel.server.enabled = true
camel.server.mcp-enabled = true
camel.server.mcp-tags = crm,notify
camel.server.mcp-server-name = my-integration-app
camel.server.mcp-server-title = My Integration MCP
camel.server.mcp-server-description = Tools for customer operations
camel.server.mcp-server-website-url = https://example.com/docs
camel.server.mcp-instructions = Use these tools to operate the integration.
camel.server.mcp-server-icons = [{"src":"https://example.com/icon.png","mimeType":"image/png","sizes":["48x48"],"theme":"light"}] Tag patterns support wildcards: use to expose all tagged tools, or a prefix pattern like crm to expose all tags starting with crm:
On other runtimes, or when wiring programmatically, add the McpServerBridge service to the CamelContext instead:
McpServerConfiguration configuration = new McpServerConfiguration();
configuration.setTags("crm,notify");
configuration.setServerName("my-integration-app");
configuration.setServerTitle("My Integration MCP");
configuration.setServerDescription("Tools for customer operations");
configuration.setServerWebsiteUrl("https://example.com/docs");
configuration.setInstructions("Use these tools to operate the integration.");
configuration.setServerIcons(List.of(new McpServerIcon("https://example.com/icon.png", "image/png", List.of("48x48"), "light")));
camelContext.addService(new McpServerBridge(configuration)); The MCP endpoint is then served at http://<host>:<port>/mcp on the Camel main HTTP server. Any MCP client can connect over streamable HTTP, for example another Camel integration using the camel-openai MCP client:
from("direct:agent")
.to("openai:chat-completion"
+ "?model={{llm.model}}"
+ "&autoToolExecution=true"
+ "&mcpServer.myCamelTools.transportType=streamableHttp"
+ "&mcpServer.myCamelTools.url=http://localhost:8080/mcp"); Options
The options, configurable as camel.server.mcp-* properties on Camel Main / JBang (see the camel-main options) or on McpServerConfiguration programmatically:
| Option | Description | Default | Owner |
|---|---|---|---|
| Whether to expose ai-tool routes as MCP tools and ai-resource routes as MCP resources over streamable HTTP. |
| bridge |
| Comma-separated list of tag patterns selecting the ai-tool and ai-resource routes to expose. Patterns support exact match, wildcard prefix ( | bridge | |
| Per-call tool execution timeout in milliseconds. A call exceeding the timeout returns an error result to the MCP client; the underlying route keeps running until it completes on its own. |
| bridge |
| Per-read resource execution timeout in milliseconds. A read exceeding the timeout returns an error to the MCP client; the underlying route keeps running until it completes on its own. |
| bridge |
| HTTP path where the MCP endpoint is served. |
| engine |
| MCP server name advertised to clients. | CamelContext name | engine |
| MCP server display title advertised to clients in | engine | |
| MCP server description advertised to clients in | engine | |
| MCP server website URL advertised to clients in | engine | |
| Top-level MCP instructions returned to clients on initialize (not part of | engine | |
| JSON array of MCP server icons advertised to clients in | engine | |
| Keep-alive ping interval in milliseconds for MCP sessions on the Vert.x streamable transport. Dead sessions are evicted after consecutive ping failures. |
| engine |
| Idle TTL in milliseconds for MCP sessions on the Vert.x streamable transport. Sessions with no activity for longer than this interval are evicted. |
| engine |
Bridge-owned options are honored identically on every runtime. Engine-owned options are consumed by the Vert.x engine only; on runtimes with a native engine (Quarkus, Spring Boot) the native configuration decides serving concerns and a startup WARN is logged when an ignored option is set.
Protocol
This section describes the Vert.x engine shipped in camel-mcp-server, which serves on Camel Main and Camel JBang. On Quarkus and Spring Boot the transport is owned by the native engine instead — quarkus-mcp-server and the Spring Boot embedded HTTP server (Spring AI MCP server) respectively — and the details below do not apply.
The Vert.x engine implements the MCP streamable HTTP transport:
-
POST /mcpansweringapplication/jsonortext/event-streamdepending on the request, -
a long-lived
GET /mcpSSE channel for server notifications, withLast-Event-IDreplay, -
session management via the
Mcp-Session-Idheader andDELETE /mcpfor session termination. -
active session eviction: keep-alive pings (default every 30 seconds) remove sessions whose ping fails repeatedly, and an idle TTL (default 5 minutes) removes sessions with no traffic. Configure with
camel.server.mcp-session-keep-alive-intervalandcamel.server.mcp-session-idle-ttl(set either to0to disable).
Tools and resources appearing or disappearing (routes starting and stopping) emit notifications/tools/list_changed and notifications/resources/list_changed to connected clients. Per-resource subscriptions (resources/subscribe) are not offered: the server advertises resources with listChanged only.
Security
External MCP clients are untrusted senders under the Camel security model. The module applies the following rules:
-
Explicit opt-in per route: only tools and resources whose tags match the configured tag patterns are exposed. The untagged default pools are never exposed, even when using the
*wildcard. -
Flat namespace protection: a tool whose name collides with an already exposed tool is refused with an ERROR log — never silently replaced. Resources follow the same rule on their uri.
-
Error sanitization: route exceptions are mapped to a generic error message; the cause is logged server-side and never sent to the client. Argument validation messages (missing or invalid parameters) are returned as-is.
-
Bounded execution: every call is subject to the
toolTimeoutand every read to theresourceTimeout. Note that a timed-out route keeps running server-side until it completes; the timeout bounds the MCP request, not the route. -
Authentication: the MCP endpoint is served through the main HTTP server router, so platform-http authentication (basic, JWT via
camel.server.authentication*options) applies to it. The MCP specification’s authorization model is OAuth 2.1; see camel-oauth for resource-server style protection. On Quarkus and Spring Boot, authentication is owned by the native runtime security.
Runtime notes
-
Camel Main / JBang: requires the Camel main HTTP server (
camel.server.enabled=truewithcamel-platform-http-main, automatic with Camel JBang) or aVertxPlatformHttpServerservice. Serving is fully asynchronous: tool calls are offloaded to the Vert.x worker pool and the long-lived SSE channel does not occupy a worker thread. -
Quarkus: use the
camel-quarkus-mcp-serverextension (serves through quarkus-mcp-server; the MCP Java SDK is not on the classpath). -
Spring Boot: use the
camel-mcp-server-starter(serves through the Spring AI MCP server).