About MCPShim

Turn any MCP server or configured HTTP API into a standard CLI command your agent can call directly.

The story

MCP servers and HTTP APIs are powerful, but using them is harder than it should be. Every MCP server has its own transport, its own auth flow, and its own way of exposing tools. An agent that wants to use tools from three different MCP servers has to manage three connections, three authentication schemes, and three different lifecycles.

Most AI agents already know how to run shell commands. They can call grep, curl, or git without thinking twice. So the question became: what if every MCP tool worked the same way - as a plain CLI command?

That's what MCPShim does. It calls remote HTTP and SSE MCP servers, binds ordinary HTTP endpoints as policy-controlled tools, centralizes authentication, and exposes each tool through one command-line interface. Your agent just runs a command and gets the result.

“One daemon. Your remote MCP servers. Standard CLI commands.”

The agent doesn't need to know it's talking to an MCP server at all. It runs a command, passes arguments, and reads the output. That's it.

How it's designed

CLI-first, always

Every MCP tool becomes a regular CLI command. Your agent calls it the same way it calls any other program on the system. Arguments go in, results come out. No special protocol, no client library required.

Auth is centralized

OAuth tokens, API keys, and header-based credentials live behind one configuration and local token store. Interactive OAuth remains an explicit setup action, including a manual headless flow.

Composable by nature

Because everything is a CLI command, you can pipe output through jq, chain tools together in a shell script, or let your agent mix MCP tools with any other command-line program. Standard Unix patterns just work.

All services, one interface

Add remote HTTP and SSE MCP servers or bind ordinary APIs as typed and constrained raw tools. Every capability uses the same discovery and calling convention.

What it's built with

MCPShim is written in Go and ships as a single binary. No runtime dependencies, no containers required. Install it and it just works.

Language

Go

Protocol

Unix socket + JSON

Storage

SQLite

License

Open source

Under the hood, a single daemon process (mcpshimd) centralizes access to every configured MCP endpoint and HTTP service. Lightweight CLI wrappers talk to the daemon over a Unix socket. Everything stays local, fast, and private.

Who it's for

MCPShim is built for AI agents and the developers who build them. If your agent needs to:

  • Use tools from remote MCP servers without managing connections
  • Expose an existing HTTP API without building an MCP server
  • Call MCP tools the same way it calls grep, curl, or any shell command
  • Centralize OAuth tokens and header-based credentials
  • Combine MCP tools with local commands in shell scripts and pipelines
  • Run headless on a server with no browser or UI required

…then MCPShim is the missing piece. Your agent focuses on reasoning. MCPShim gives it the tools.

Built with care

MCPShim is an open source project by CBK AI. You can explore the code, report issues, or contribute on GitHub.

Also check out Pantalk - a sibling project that gives AI agents a unified interface to chat apps like Slack, Discord, WhatsApp, and Telegram.