MCP tools

Every agent session gets a small MCP server named micromanager, with tools to ask you a question, read the workspace’s diff, leave and read review comments, read terminal output and look up the workspace. The tools Conductor had are here under the same names, so prompts and habits that mention them carry over; GetWorkspaceInfo is microManager’s own.

Updated

Which tools does the micromanager server offer?

ToolInputWhat it does
AskUserQuestion { questions: { question, header?, options: { label, description? }[], multiSelect? }[] } Shows a question card in the chat and returns your answer. Offered to non-Claude agents; Claude Code uses its own built-in tool of the same name, which microManager renders the same way.
GetWorkspaceDiff { path?: string, stat?: boolean } Returns the workspace’s unified diff against the merge base with its target branch, including uncommitted and untracked files. Pass a path for one file, or stat for a summary.
DiffComment { comments: { file, line, endLine?, body }[] } Adds review comments from the agent to the Review panel, anchored to lines of the diff.
GetDiffComments { file?: string } Returns the review comments on the diff that are not resolved yet, yours and agents’, each with its file, lines and the code it is on, so “address my comments” works without sending them. Pass a file for one file’s. Reading them changes nothing.
GetTerminalOutput { terminal?: "run" | "setup" | <terminal id>, lines?: number } Reads the last lines of output from the run script (the default), the setup script or a terminal in the workspace.
GetWorkspaceInfo {} Returns the workspace’s name, repository, folder, main checkout, branch, target branch, reserved ports, setup status and pull request.

What does a DiffComment call look like?

Comments appear in the workspace’s Review panel next to yours, marked as coming from the agent. You can resolve them, or send the open ones to the chat you are in.

{
  "comments": [
    {
      "file": "src/api/users.ts",
      "line": 42,
      "endLine": 48,
      "body": "This retries forever when the token is revoked; cap it at three attempts."
    }
  ]
}

How do agents connect to it?

  • Claude Code gets the server in-process, through the Claude Agent SDK.
  • ACP agents that support HTTP MCP servers reach it over Streamable HTTP on 127.0.0.1, at a random port, with a bearer token made for that one session.
  • Codex, and ACP agents that only speak stdio MCP, get a small bridge that forwards to that same endpoint, with the token in its environment rather than on its command line.

Can other programs or web pages call these tools?

No. The endpoint listens only on 127.0.0.1, every request needs that session’s 256-bit token, and requests carrying an Origin header, which browsers add to requests web pages make, are refused. It is the only network port microManager opens; the app’s own interface talks to its backend in-process, not over HTTP.

Do my own MCP servers still work?

Yes. The micromanager server is added alongside the ones you configured for each agent. Claude Code, for example, still loads the MCP servers from your user settings, and from the repository’s .mcp.json and project settings once you trust the repository.

Where to go next

Get microManager

Free, no account, for macOS, Windows and Linux. It runs the agent CLIs you already have installed and signed in.

curl -fsSL https://micromanager.win/install.sh | sh

That installs it on macOS and Linux. On Windows, paste this into PowerShell:

irm https://micromanager.win/install.ps1 | iex