Connect an agent
Every agent needs its own Plow credential. Your hosts recognize the agent by it, and judge each of its requests as that agent's: against your rules, and what you told it in your Plow chat.
Never copy your own ~/.latch folder, or its token, to an agent. A host would take the agent's requests for yours.
An agent Plow runs
Agents built on Plow's OpenClaw or Hermes images: nothing to set up. latch comes installed and signed in.
Any other agent on Plow: its credential already works with latch. Give it the Agent install, then latch login.
Your own agent
An agent you run yourself, anywhere: your laptop, a server or a container. It gets its credential in three steps.
1. Pick a phone line. Each agent answers on one of your Plow phone lines. That's where you chat with it. From one of your own computers:
latch agents lines2. Register the agent. This tells Plow about an agent you run yourself (that's what --own means), on that line, and writes its credential to a new file that only you can read:
latch agents new --own --name my-agent --line LINE --token-file my-agent.tokenLINE is a line's number from latch agents lines. Treat my-agent.token like a password.
3. Give the agent its credential. Copy the file to the agent's computer. There, choose the agent's state folder before the agent install, then install and log in:
export PLOW_TOKEN_FILE=~/my-agent.token
export LATCH_STATE_DIR=~/latch-agent
curl -fsSLO https://releases.plow.co/latch-cli/install.sh && sh install.sh --client-only
latch loginlatch reads the agent's credential from the file PLOW_TOKEN_FILE names, or from PLOW_AGENT_TOKEN if you set the token itself. Keep them set for every command the agent runs: latch never copies the credential into its state folder. Keep LATCH_STATE_DIR set too: the install marks that folder as the agent's own computer, which latch update checks.
Use a state folder of its own. A folder that already holds your own login uses yours, whatever the agent's credential says.
To revoke the agent, run latch agents remove my-agent.
What an agent may do
By default, agents may look things up, open web pages, work on your files, run project builds and tests, and install packages the project already lists. Sending, inviting, deleting other files and changing accounts or shared systems need your permission. Buying, reading secrets or sharing private data must be named explicitly in your rules or your own Plow chat. A task you give on a site includes using its saved login there; it does not extend to buying, sending or changing things after sign-in.
A permission adds to these defaults. A rule that grants one task leaves other everyday work allowed; a restriction still holds. “Never” stays absolute. “Unless I ask in chat” lets you approve the action yourself, in your own messages. An agent quoting you cannot give permission.
For example:
- “Do not read or search this Mac's Messages.”
- “Agents may email the people in my contacts.”
- “Never buy anything unless I ask in chat.”
Each host keeps its own rules: see How Latch works.
Built for agents
Every command prints one JSON object, and errors print as JSON too.
latch skillprints the guide written for agents.latch help COMMAND --examplesshows how to call a command.--host NAME, orLATCH_HOST, picks the host. See Choosing a host.
latch exec --host my-mac -- ls -la| Exit code | Means | Do |
|---|---|---|
| 0 | Done. | |
| 1 | Failed. | Read the error. |
| 2 | The command is wrong, or it names no host. | Fix the command. |
| 3 | The host couldn't be reached, or couldn't judge the request. | Retry later. A browser plan may have run some steps: read its steps. |
| 4 | Refused. | Don't retry. For not_covered, ask the owner in your Plow chat, then retry. |
| 5 | Sent, but no answer came back. | Check what happened before you send it again. |
This page is made from its source in latch-cli. Found a mistake? Open a pull request there.