The archived Go CLI that taught LLMs to use external tools

What it does

MCPHost acted as a bridge between LLMs and external capabilities. It managed connections to MCP servers—local processes, remote HTTP endpoints, or built-in in-process modules like filesystem, bash, and HTTP fetchers—so models from Anthropic, OpenAI, Google, or Ollama could discover and invoke tools. The toolset was configurable per server, with allowlists and blocklists to restrict what the model could actually touch.

The interesting bit

The project treated tool access as a security boundary rather than an afterthought. Its environment-variable substitution syntax let operators inject secrets and settings into YAML configs without hardcoding them, and a hooks system offered custom integration points for policy enforcement.

Key highlights

• Supports local, remote (StreamableHTTP), and built-in MCP server types concurrently
• Built-in servers include filesystem, bash execution, ephemeral todo lists, and HTTP fetching
• Tool filtering via allowedTools and excludedTools per server
• Go SDK for programmatic use without spawning CLI processes
• Environment variable substitution with defaults in configuration files

Caveats

• No longer maintained: the repository is archived and superseded by Kit
• The README warns that TLS certificate verification can be disabled with --tls-skip-verify, which the authors flag as insecure for production

Verdict

Worth studying if you are building an MCP client or curious about the protocol’s early CLI tooling, but active projects should migrate to Kit instead.