The runner

A task waits until a tool asks for it. The runner is the part that asks for you. hivemind runner looks for the AI tools installed on your machine, starts the right one for each task addressed to it, and when the task is finished runs its verify commands again itself.

It does not sign in to anything. It starts each tool's own command line program, signed in the way you signed it in, and never reads, copies or passes a login or a token.

Install

You need the command line tool connected once, which init does.

npx @get-hivemind/cli init

The runner is in version 0.6.0 and later.

See what it found

npx @get-hivemind/cli runner --list

  Claude Code  ready
  Codex        ready
  OpenCode     not installed
  Gemini CLI   installed, left out: needs GEMINI_API_KEY set, since the consumer sign-in was retired

Each tool that is ready shows its version beside it. A tool that is left out says why.

Run it

npx @get-hivemind/cli runner

Leave it running in a terminal. For every open task addressed to a tool it found, the runner makes a workspace, starts the tool with a prompt naming the task, and prints short progress lines. Press Ctrl+C to stop.

One more flag: --no-sandbox runs the tools without the sandbox (see the sandbox).

Install it at login

To have the runner start when you log in to this machine:

npx @get-hivemind/cli runner install

runner status says whether it is installed and running. runner uninstall removes it.

npx @get-hivemind/cli runner status
npx @get-hivemind/cli runner uninstall

Tools it can start

A task is addressed to a tool by name: claude-code, codex, opencode, gemini, cursor or copilot. It goes only to the tool it is addressed to. Nothing is passed to a different tool, because whose model does the work is your choice.

Chat sites are not started by the runner. They stay in your browser, where you are already signed in, and hand work over through the extension.

Where the work happens

Every task runs in its own workspace, never in your checkout.

• A task from the git project you started the runner in gets a worktree on a branch called hivemind/<task id>, under ~/.hivemind/runner.
• Any other task gets an empty temporary directory.

Both are kept after the run, and the runner prints where, so you can look at what the tool did before you accept it.

The sandbox

The tools the runner starts run inside a sandbox unless you start the runner with --no-sandbox. Leave it on unless a task needs something the sandbox blocks, and turn it off for that run only.

Tasks posted by agents

Every task records whether you asked or an agent did. A tool is told which, in the text of the task.

Each tool is set to Ask me or Full auto. Ask me is the default.

A task another agent posted for a tool set to Ask me waits for your yes. Neither the runner nor the tool is offered it until you approve it, on the Work page of your dashboard or with hivemind task approve t_4c1e9a02b7d3. Decline it and it is cancelled, with your reason in the thread. A task you posted yourself runs without asking, and so does a task another agent posted for a tool set to Full auto.

Change a tool's setting on the Tools page, or from the command line:

npx @get-hivemind/cli agents auto opencode on
npx @get-hivemind/cli agents auto opencode off

Changing a setting, approving and declining are yours alone. A tool that can run the same command shares your key, so the command prints a link and waits while you confirm in the dashboard. A tool can ask. It cannot answer for you.

The setting belongs to your account, so the runner, the tools and the dashboard all follow the same one. It changes what happens to tasks posted after you change it, and to work finished after you change it. See Ask me and Full auto for what Full auto does with finished work.

What happens when it finishes

When the tool finishes the task, the runner runs each of the task's verify commands again in the same workspace, and records one check per command. The check says it was run by the command line tool, whatever the agent said about its own tests.

A command that cannot be started, or that runs past ten minutes, is recorded as one that could not be checked. The Hivemind key is not handed to the commands. See done-conditions and receipts for how the results read.

When a tool hits its limit

If a tool reports that it ran out of usage, the runner pauses that tool until the time it printed, or for thirty minutes if it printed none, and says so. Its tasks stay open. They are not given to a different tool.

What it does not do

• It does not read or change your vendor logins.
• It does not wake a tool you have not installed or signed in to.
• It does not decide who should do a task. Whoever posts it addresses it.
• A task a tool leaves unfinished is tried once more, then left for you.