Skip to main content

​
3. Native MCP Integration

Status: Accepted Date: 2026-02-09

​
Context

crocbot used a subprocess-based mcporter CLI skill to invoke MCP tools. Each tool call spawned a new process, adding ~200ms overhead and preventing streaming responses. The MCP specification matured enough to warrant a first-class integration.

​
Decision

Build an in-process MCP client and server using @modelcontextprotocol/sdk:
  • Client: McpClientManager with lifecycle management (connect, getTools, callTool, disconnect, shutdown). Supports stdio (local), SSE (remote), and streamable HTTP transports. All remote transports pass SSRF validation.
  • Server: Expose crocbot as MCP infrastructure via SSE and streamable HTTP transports. Bearer token authentication. Four tools: send_message, finish_chat, query_memory, list_capabilities.
  • Tool bridge: Convert MCP tool schemas to agent tools with collision detection against existing tool names.

​
Consequences

​
Enables

  • Zero-overhead tool invocation (no subprocess spawn)
  • Streaming responses from MCP tools
  • External AI systems can call into crocbot via MCP server mode
  • SSRF-protected remote connections

​
Prevents

  • Using the old mcporter subprocess approach (marked READY FOR REPLACEMENT)
  • Unguarded connections to remote MCP servers

​
Trade-offs

  • Adds @modelcontextprotocol/sdk as a runtime dependency
  • MCP server mode requires a shared token for authentication
  • Protocol drift risk as MCP spec evolves (mitigated by official SDK)