26 Unix tools. One binary. Zero dependencies. · the missing coreutils for AI pipelines · vrk mcp - expose all 26 tools to any AI agent · brew install vrk - ready in 5 seconds · 26 Unix tools. One binary. Zero dependencies. · the missing coreutils for AI pipelines · vrk mcp - expose all 26 tools to any AI agent · brew install vrk - ready in 5 seconds

MCP

MCP server

vrksh includes a discovery-only MCP server. It exposes all pipeline tools via the Model Context Protocol so agent frameworks can discover what tools are available.

The MCP layer handles discovery, not execution. Tools run as shell commands - the MCP server tells your agent what exists and what flags each tool accepts.

Claude Code

Add to your MCP config:

{
  "mcpServers": {
    "vrksh": {
      "command": "vrk",
      "args": ["mcp"]
    }
  }
}

Every tool becomes visible in Claude Code’s tool list with its full input schema.

How it works

vrk mcp starts a JSON-RPC 2.0 server over stdio. It responds to:

  • initialize - returns server capabilities
  • tools/list - returns all tool definitions with input schemas

tools/call is intentionally not implemented. The agent executes tools via shell:

echo "hello" | vrk tok --json

This keeps the execution model simple and auditable - every tool invocation is a shell command that appears in your agent’s logs.

Alternative: --manifest

If you don’t need MCP, use vrk --manifest for a JSON tool registry:

vrk --manifest | jq '.tools[].name'

Or vrk --skills for the full reference with flags, exit codes, and gotchas:

vrk --skills tok

See the agent endpoints page for a complete index of all machine-readable surfaces.