Skip to main content

MCP Ecosystem Reference

Model Context Protocol (MCP) Ecosystem​

The Cortex workspace houses the containerized MCP ecosystem. It exposes structured development tools and codebase contexts to LLM clients (such as the Antigravity CLI and GitHub Copilot CLI) through a single unified ingress proxy gateway.

1. Unified Gateway Routing​

Instead of coupling client agents directly to individual MCP tool servers, all tool interactions are federated through the Go-based AgentGateway proxy.

  • Unified Ingress Endpoint: The gateway listens on port 8080. Inside the container bridge network (tupynambalucas-cortex-net), agents connect to a single endpoint: http://agentgateway:8080/mcp/http.
  • Downstream Federation: The gateway acts as a reverse proxy. It intercepts tool calls and delegates them to target containerized MCP servers running in the same network namespace (mcp-github, mcp-context7, mcp-browser).
  • Host Connectivity: Development tools running on the host machine loopback interface communicate with the gateway using http://localhost:8080/mcp/http.

2. Zero-Dependency HTTP/SSE Adapters​

Standard input/output (stdio) based MCP servers are wrapped in custom TypeScript adapters (sse-adapter.ts) running on Fastify. This adapter layer translates stdio streams into standard web protocols:

  • Traditional Server-Sent Events (SSE): Establishes a uni-directional stream via GET /sse and accepts client messages via POST /messages, piping events directly to the underlying subprocess stdin.
  • Streamable Stateless HTTP (POST /sse): Simulates stateless request-response by buffering stdout lines from the child process, extracting JSON-RPC payloads, and resolving the HTTP request only when it captures the specific JSON-RPC response id matching the request.

3. Integrated Context Servers​

A. GitHub MCP Server​

Connects the containerized development environment directly to the GitHub API, enabling AI agents to execute Git actions, coordinate branch states, and manage issues.

  • Tools Registry:
    • create_branch, list_branches, update_pull_request_branch
    • get_commit, list_commits, get_tag, list_tags
    • get_file_contents, create_or_update_file, delete_file, push_files
    • issue_read, issue_write, list_issues, add_issue_comment
    • pull_request_read, create_pull_request, update_pull_request, list_pull_requests, merge_pull_request

B. Context7 MCP Server​

Queries external documentation APIs in real time to fetch syntax specifications and usage guides for core dependencies (React 19, Fastify 5, Three.js) without relying on stale model training data.

  • Tools Registry:
    • resolve-library-id: Maps a package name to its unique Context7 ID.
    • query-docs: Retrieves updated API sections, guides, and code examples.

C. Browser MCP Server (Playwright)​

A headless Google Chrome instance managed via Playwright, enabling agents to navigate pages, capture screenshots, audit console errors, and perform visual regression tests.

  • Loopback Routing: Rewrites destination URLs targeting localhost or 127.0.0.1 to host.docker.internal to route requests successfully from the container bridge back to the host machine.
  • Tools Registry:
    • browser_navigate, browser_click, browser_type, browser_fill_form, browser_select_option
    • browser_take_screenshot, browser_snapshot, browser_console_messages, browser_network_requests
    • browser_evaluate, browser_resize, browser_close

D. Docker Hub MCP Server​

Manages interactions with the Docker Hub registry to inspect container metadata and track image tags.

  • Tools Registry:
    • search, getRepositoryInfo, checkRepository, listRepositoryTags, getRepositoryTag