Model Context Protocol (MCP)
PROTOCOLThe Model Context Protocol (MCP) standardizes how AI models connect to external tools and live user interfaces. Firebreak implements an official MCP server conforming to the protocol specification.
4 min read
•Production Verified
Universal Dual-Transport Architecture
Firebreak's MCP server operates across two distinct transport modes:
| Transport | Protocol | Primary Consumer | Port / Pipe |
|---|---|---|---|
| STDIO | JSON-RPC 2.0 over Standard I/O | Local desktop AI agents (Claude Desktop, Cursor CLI) | Process stdin / stdout |
| HTTP SSE | JSON-RPC 2.0 over Server-Sent Events | Cloud-hosted AI models, NitroStack Studio, web agents | Port 3001/3005 (/sse) |
MCP Server Architecture: Dual Transports (STDIO & SSE), McpKeyGuard, and Tools / Widgets Pipeline
Scoped Authentication (McpKeyGuard)
Cloud requests must supply an Authorization: Bearer fb_mcp_... header. McpKeyGuard hashes the key and asserts required scopes:
| Scope | Permissions |
|---|---|
| services:read | Allowed to discover protected services (list_services). |
| incidents:read | Allowed to list incidents, inspect suspects, and retrieve timelines. |
| actions:propose | Allowed to stage rollback proposals (propose_rollback). |
| actions:approve | Allowed to trigger live production rollbacks (approve_action). |