Skip to content
alfe.aiDocs

Use MCP servers with your agent

Any Model Context Protocol server you attach to an agent becomes part of that agent’s toolset. The server’s tools show up alongside the agent’s built-in tools, and the agent calls them in the normal course of its work.

An agent doesn’t need to know whether a tool comes from its runtime, from an Alfe integration, or from an MCP server you wired up yourself.

How an agent hosts MCP servers

Section titled “How an agent hosts MCP servers”

Each agent runs a small MCP host as part of its local runtime. When you attach one or more MCP servers to the agent, the host:

  • Keeps long-lived connections to each server, so tools are ready without a cold start on every turn.
  • Presents one unified toolset. Tools from every attached server are merged into a single catalog the agent sees. Names are namespaced per server, and collisions between servers are disambiguated automatically, so two servers can both expose a search tool without clashing.
  • Routes tool calls back to the server that owns each tool, and returns the result to the agent.
  • Reacts to changes. It picks up added and removed servers on its own, so you don’t restart the agent’s runtime.

Because the host runs alongside the agent, the servers you attach are scoped to that agent — attaching a server to one agent never exposes its tools to another.

Types of MCP servers

Section titled “Types of MCP servers”

Two broad categories, both consumed the same way once attached:

Category What it is How you add it
Integration-provided servers Curated MCP servers that ship as Alfe integrations — e.g. app and productivity tools. Installing the integration wires the server up for you, including credentials from your connected account. Dashboard, CLI, or the agent itself. See Add an MCP server.
Your own servers Any MCP server you want to attach directly — a local command-line server on the agent’s machine, or a remote server reachable over HTTP. The alfe mcp commands.

Integration-provided servers are the durable, credential-aware path and are managed through the integrations lifecycle. Your own servers are the escape hatch for anything not in the catalog. Both are covered in Add an MCP server to your agent.

The built-in Alfe tools

Section titled “The built-in Alfe tools”

Agents also get Alfe’s own platform tools through an MCP server called alfe-platform. What it offers depends on the runtime:

  • OpenClaw gets a small set, because its native Alfe plugins provide the rest.
  • Claude Code, Codex, Gemini CLI, and Grok Build get more, because they have no plugins. That includes memory, voice, messaging, search, secrets, shared knowledge, webhooks, and scheduled jobs, plus the browser tools on runtimes that support them.

Transports

Section titled “Transports”

An agent can consume MCP servers over the standard transports:

  • stdio — a local server the agent launches as a child process (a command with arguments and environment variables).
  • Streamable HTTP and SSE — remote servers the agent connects to over HTTP, optionally with request headers (for example, an authorization token).

Seeing what tools a server exposes

Section titled “Seeing what tools a server exposes”

To check which tools a registered server advertises — useful when a server is attached but its tools aren’t showing up — connect to it briefly and list them:

Terminal window
alfe mcp test <id>

This connects to the server, prints the tools it advertises (with the namespaced names the agent will see), and disconnects. See the alfe mcp reference for the full command set.

Next steps

Section titled “Next steps”