Skip to content

Learn · Track

MCP servers with an agent

How to run the open-source MCP servers people actually use under an agent, one server per guide: what it does, the version-pinned command from its README, how to attach it, which of its tools write, and how to keep it narrow with read-only tools, an allow-list and approvals for writes. Each guide ends with a way to test it without a model, and says where a reference server has been archived.

Guides in this track

Intermediate10 guides, in reading order

  1. mcpServers JSON config

    What each field in an MCP client config does (command, args, env, url, headers, inputs) and how to move servers between clients without leaking secrets.

    8 min read

  2. Filesystem MCP server

    Run the reference filesystem MCP server for an agent: allowed directories, read-only use, write approvals and the path checks that keep it contained.

    8 min read

  3. GitHub MCP server

    Connect the open-source GitHub MCP server to an agent with a fine-grained token, read-only toolsets, and approvals before issues, PRs or merges.

    10 min read

  4. Read-only SQL over MCP

    Connect a database to an agent through MCP with a read-only role, statement timeouts, row limits and approvals, so a prompt can't drop a table.

    11 min read

  5. Fetch MCP server

    Use the reference fetch MCP server with an agent: what it returns, robots and size limits, and why untrusted pages need SSRF and injection controls.

    9 min read

  6. Git MCP server

    Run the reference git MCP server under an agent: read-only tools first, which tools write, and how to keep an agent from committing without approval.

    7 min read

  7. Playwright MCP server

    Run the open-source Playwright MCP server for an agent: accessibility snapshots vs screenshots, isolated profiles, and limits on where it can browse.

    8 min read

  8. Memory MCP server

    What the reference memory MCP server (a knowledge graph) stores, how it differs from session memory in an agent runtime, and when to use each.

    8 min read

  9. Time and everything servers

    Use the reference "everything" and "time" MCP servers to exercise a client: tools, resources, prompts, progress and errors, without credentials.

    8 min read

  10. Build a minimal MCP server

    A 40-line MCP server with one tool, run over stdio and Streamable HTTP, then tested with recorded calls so CI can check it without a model.

    8 min read

Before and after

Prerequisites: What is MCP? and MCP transports, so the difference between a local command and a remote URL is clear.

After this track: Least privilege for agents turns the per-server advice into a method, and Test MCP servers in CI shows how to keep a server's tools from changing under you. To try an agent with no API key, use the zero-key MCP agent recipe.

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings