MCP — Quick start (consume)
llmist’s CLI can connect to Model Context Protocol servers and merge their tools into the agent’s gadget catalog. Use --mcp-server for ad-hoc stdio servers, or persist stdio and Streamable HTTP servers in [mcp.servers.<name>] TOML config blocks.
5-minute walkthrough
Section titled “5-minute walkthrough”-
Set an API key (any one of these):
Terminal window export OPENAI_API_KEY="sk-..."# or ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENROUTER_API_KEY -
Run an agent with the public Filesystem MCP server:
Terminal window llmist agent \--mcp-server fs="npx -- -y @modelcontextprotocol/server-filesystem /tmp" \"list files in /tmp and tell me how many there are"
The agent connects, discovers list_directory, read_file, etc., picks the right tool, and returns the count. The MCP child process is terminated when the agent exits.
The --mcp-server flag
Section titled “The --mcp-server flag”--mcp-server <name>=<command> [-- <args>...]Format details:
<name>— a stable identifier for this server. Surfaced in logs.<command>— the executable. Its basename must be in the default allowlist (npx,node,uvx,python,python3,deno,bun) — or the server must be marked trusted via--mcp-trust <name>.<args>— anything after--is passed as arguments to the executable. (No--and a multi-token tail also works; the first token is the command, the rest are args.)
Repeat --mcp-server to attach multiple servers. If tool names collide across servers, llmist deterministically prefixes the affected tools with <server>__.
llmist agent \ --mcp-server fs="npx -- -y @modelcontextprotocol/server-filesystem /tmp" \ --mcp-server git="npx -- -y @modelcontextprotocol/server-git" \ "show me the most recently modified file in /tmp and its git history"The --mcp-trust flag
Section titled “The --mcp-trust flag”If your server’s command isn’t in the default allowlist:
llmist agent \ --mcp-server custom="/Users/me/bin/my-mcp -- --port 1234" \ --mcp-trust custom \ "use the custom server to do X"This bypasses the runtime allowlist for that one server. Read MCP security before you reach for it.
TOML configuration
Section titled “TOML configuration”For repeat usage, put servers in your llmist config:
[mcp.servers.fs]transport = "stdio"command = "npx"args = ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]timeout-ms = 30000
[mcp.servers.remote]transport = "http"url = "https://my-mcp.example.com/mcp"
[mcp.servers.remote.headers]Authorization = "Bearer xyz"Invalid [mcp] config fails fast with a path-specific error. timeout-ms is a non-negative integer; 0 disables the timeout.
If you already have MCP servers configured in Claude Code, import them:
llmist mcp import-claude-code >> ~/.llmist/config.tomlSee TOML configuration — MCP servers for the full schema.
Prompts and expose side
Section titled “Prompts and expose side”Servers that advertise MCP prompts expose them as slash-invocable llmist skills. To publish llmist gadgets and skills for other MCP clients, use llmist mcp serve.
Library API
Section titled “Library API”The same machinery is available in code:
import { LLMist } from "llmist";
const answer = await LLMist.createAgent() .withModel("sonnet") .withMcpServer({ name: "filesystem", transport: "stdio", command: "npx", args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"], }) .askAndCollect("list files in /tmp");See MCP library reference for the full type surface.