Skip to main content

​
Usage

toolshed serve
Starts an MCP server that communicates over stdio, allowing MCP clients like Claude or Cursor to connect.

​
Prerequisites

You must be logged in first. Run toolshed login if you haven’t already. The command will exit with an error if no config is found.

​
How it works

  1. Loads credentials from ~/.toolshed/config.json
  2. Fetches the tool catalog from GET /api/registry using the session token
  3. Creates a runtime (local or Vercel sandbox) with the user’s role and tool set
  4. Starts the MCP server on stdio transport
  5. When an agent calls a tool, the CLI vends a token from POST /api/tokens/vend and executes the tool handler

​
Connecting from Claude

Add Toolshed to Claude’s MCP server config: 
{
  "mcpServers": {
    "toolshed": {
      "command": "node",
      "args": ["<path-to-repo>/apps/cli/dist/index.js", "serve"]
    }
  }
}
Claude spawns the process, connects over stdin/stdout, and discovers tools via MCP’s tools/list.

​
Environment variables

VariableDefaultDescription
TOOLSHED_RUNTIMElocalRuntime backend: local or vercel
VERCEL_TEAM_ID—Required when TOOLSHED_RUNTIME=vercel
VERCEL_PROJECT_ID—Required when TOOLSHED_RUNTIME=vercel
VERCEL_TOKEN—Required when TOOLSHED_RUNTIME=vercel

​
Authentication flow

When an agent calls a tool that requires OAuth credentials (e.g., github.issues.create):
  1. The CLI looks up the tool’s authProvider field
  2. Calls POST /api/tokens/vend with the user’s ID and provider name
  3. The server decrypts the stored OAuth token (refreshing if expired)
  4. The CLI receives a usable access token and passes it to the tool handler
Tokens are vended per-call — they never persist on disk or in the MCP transport.