Autonomous Harness MCP Server: Let Your Assistant Start and Watch Other Coding Agents

• Shows who you are to Harness

  Says which agent is calling, on which machine, whether it may start harnesses now, how many it has open and how many starts are left. It is meant to be called first.

  Try: “Check whether you can start Harness agents.”

  Read · harness_whoami

• Lists running agents

  Lists the harnesses on a machine with each one's agent, state, who started it, last activity, folder and topic. It can include ones stopped in the last week.

  Try: “List my Harness agents on the GPU box.”

  Read · harness_list

• Searches past conversations

  Searches every turn of every live or stopped harness on this computer for words, so you can find what was already said or decided before starting new work.

  Try: “Search Harness for what we decided about the database.”

  Read · harness_search

• Lists your machines

  Lists the machines on your Harness account with their status, so a harness can be started on another one.

  Try: “Which of my machines are online?”

  Network · harness_machines

• Starts a coding agent

  Starts an agent such as Claude Code, Codex or OpenCode on a task, in a new project folder Harness makes or in an existing folder, here or on another machine. You can watch it in the Harness app.

  Try: “Start a Codex agent named reviewer to review the auth module.”

  Runs code · harness_start

• Waits for agents

  Waits until the named harnesses stop working and reports each one's state. After the timeout it says they are still working so you can call it again.

  Try: “Wait for the reviewer agent.”

  Read · harness_wait

• Reads an agent answer

  Returns a harness's state and its last turn: what it was asked and what it answered.

  Try: “Show me the reviewer's answer.”

  Read · harness_read

• Shows or changes agent permissions

  Lists what agents may do in Harness. With a name it turns a setting on or off. Turning one on is meant to follow your yes.

  Try: “Show which agent settings are on in Harness.”

  Write · harness_settings

• Closes an agent you started

  Closes a harness this assistant started once it is done. The conversation is kept and you can resume it. It refuses while the agent works unless told to close now.

  Try: “Close the reviewer agent.”

  Write · harness_close

Watch out for

• It starts agents that run code on your machines

  harness_start launches Claude Code, Codex or another agent on a task, here or on another machine. Those agents can edit files and run commands, so rely on their own permission prompts. github.com

• Every agent it starts spends your tokens

  Started agents use your own agent plans. The server tells the assistant to start one only for work that runs in parallel or needs another view, and allows at most 4 at once. github.com

• harness_settings can widen what agents may do

  An agent turns a setting on by passing a flag that says you agreed. That flag comes from the assistant, so make sure it asks you first. github.com

• Setup pipes a remote script into your shell

  The Harness install command is curl ... | bash. The script (cli/scripts/install.sh) downloads a checksum-verified Node runtime into ~/.harness/runtime, adds ~/.local/bin to your shell startup file, and on Linux uses apt-get, which asks for your password. Read the script first if you do not trust the host. github.com

• Starting needs a setting you turn on

  The setup text says agents may start others only after you turn on "Agents can start other agents" in Harness settings. Leave it off if you do not want that. github.com

Security advisories