An MCP server to define tools as short-lived containers. Each tool call is run within a separate Linux container.

Do you want your AI app to use a tool for which there isn't an MCP server yet? Do you want to prototype or experiment with custom tools? Then this MCP server is for you.

Installation

Build from source

go install github.com/mgoltzsche/tool-containers-mcp/cmd/tool-containers-mcp@latest

Usage

With Claude

1. Define your tools within /etc/tool-containers-mcp/tools.yaml as within the example.
2. Configure your ~/.claude.json as follows:

{
  "mcpServers": {
    "tool-containers": {
      "command": "tool-containers-mcp",
      "args": [
        "--config=/etc/tool-containers-mcp/tools.yaml"
      ]
    }
  }
}

Via HTTP

MCP can also be served via HTTP by specifying e.g. --address=:9090. There are two HTTP endpoints, one for each transport variant:

• /sse - MCP via SSE.
• /mcp - MCP via streamable HTTP.

Development

To build the binary for the host architecture using Go, run:

make tool-containers-mcp

Run the tests:

make test

Run the linter:

make lint

Build a snapshot release (without publishing it):

make snapshot

Run the docker compose example (using LocalAI for inference and MCP-Bridge as agent):

make compose

Run an example inference query against MCP-Bridge's OpenAI-compatible Chat Completion API (within another terminal):

curl -fsS http://localhost:9000/v1/chat/completions -H "Content-Type: application/json" -d '{
	"model": "qwen3-4b",
	"messages": [
		{"role": "user", "content": "Which tools do you have access to?"}
	]
}' | jq .