# From your coding agent

At the end, your coding agent plans a run with you, starts it on your go and brings you every gate, without you leaving the harness you are in.

## At a glance

**Time:** 2 minutes.

**You need:**

- [On your laptop](laptop.md) done
- Claude Code, Codex or pi, open in a Git repository of yours

## Steps

1. **Put the `loopgate` command on your path.**

   ```sh
   npm install -g loopgate
   ```

   You should see: `loopgate --version` print a version.

2. **Install the skill.**

   ```sh
   loopgate skill install
   ```

   You should see: the four folders it wrote, `~/.agents/skills/loopgate` first, then the links to it under `~/.claude/skills`, `~/.codex/skills` and `~/.pi/agent/skills`.

3. **In your harness, say what you want run**, such as `run plan-build-review on this, Fable plans and Sol builds`. Name models you are signed in to.

   You should see: the workflow drawn in a code block, then the plan: every step's engine, access and pace, the rounds, and the gates you will answer.

4. **Say yes.**

   You should see: the run's link. It opens the same run in the workbench.

5. **Keep working.** The agent waits with `loopgate wait` and relays the ask when the run needs you.

   You should see: the step's reply and the question in your harness, in words.

6. **Answer in your own words:** an option's number or name (such as `Open PR`), Continue, Lock, yes or no.

   You should see: the agent run `loopgate decide` with your words, and the run go on.

## Good to know

- Your agent uses six commands: `loopgate workflows`, `plan`, `run`, `wait`, `decide` and `tell`. [The commands](commands.md) lists the rest.
- A gate is yours. The agent asks you and passes on your answer; it never answers one itself.
- One task is one run. Parallel work is lanes inside it, never several runs.
- `tell` carries your words to a step verbatim: a reply while it waits, a steer while it works, a note before it runs.
- `loopgate plan … --on <runner>` and `loopgate run launch.yaml --on <runner>` start on your team's runner, in its workspace with the same Git origin.
- Nothing costs a token between those moments. `loopgate status <run>` and the link show where it is.

## If something is off

- **Codex cannot reach Loopgate.** Its sandbox blocks the daemon's socket. Let it run `loopgate` escalated, or allow it in `~/.codex/config.toml`:

  ```toml
  [sandbox_workspace_write]
  network_access = true
  ```

- **`skill install` says a different skill exists.** Another skill sits in one of the four folders. Move it aside and run it again.
- **The plan names another engine than yours.** The model you named is not signed in, so a ready one was fitted in its place. Sign in to it, or tell the agent which signed-in model to use.

## Next

- [The commands](commands.md)
- [The patterns](patterns.md)