Skip to main content
Claude runs on your computer, using your own Claude sign-in. AgentVault’s Claude connector (called a bridge) links it to the app, keeps it running in the background, and lets you decide what it can do. It’s the framework to pick if you want an agent that can do real work on its computer, with your permission.

Before you start

  • Signed in to Claude on the agent’s computer. The bridge uses that sign-in, so your Claude subscription covers your agent.
  • Node.js 18 or newer. In a terminal, node --version should print v18 or higher.
  • macOS or Linux. The background service uses launchd on macOS and systemd on Linux.

Connect it

1

Build the agent in the app

Open the Agents tab and tap Build a new agent. Give it a Hub Name and tap Continue.On the Intelligence step, tap Anthropic Claude · Native. Tap Continue, then Save & Connect.
2

Copy the command

You’ll see “<your agent> created!” and a command. Tap Copy Command. It looks like this:
<name> is your agent’s name in lowercase. <token> is a one-time connection code that lasts 24 hours. Don’t share it.
3

Run it on your agent's computer

Open a terminal on that computer, paste the command, and press Enter.It installs the bridge, sets it up as a background service, and hands your terminal back. You’ll see “now APPROVE the device in AgentVault — the service is waiting to enrol.”
4

Approve it

In the app, you’ll see “<your agent> enrolled!” with a short code. If your terminal shows a fingerprint, make sure it matches. Then tap Approve & Chat.If anything looks off, tap Reject.
5

Say hello

Approving opens a chat. Send a message. When Claude answers, you’re done.

How you’ll know it worked

The card on the Agents tab shows a green dot, Connected, and the Claude badge.

Keep it running

You don’t need to do anything. The command already installed a background service. Your agent keeps running after you close the terminal, and it starts again after a restart.

Check on it

Run these on the agent’s computer. Each one needs your agent’s name in front. The logs live in ~/.agentvault/claude-room-bridge/<name>/logs/. There are two files: bridge.log and bridge.error.log. The bridge writes its own messages, including errors, to bridge.error.log, so start there. To watch it live, run tail -f on that file.
To restart, use restart, never install. Running install again rebuilds the service from your current terminal. Anything special you set the first time can be lost.

Update it

1

Install the new version

2

Restart the service

You’ll see “restarted service…” Your agent comes back with the same name, rooms and history.

Tools and permissions

A new Claude agent can only chat. It can’t run commands or touch files until you say so. You control that with two switches on the agent’s card, on the Agents tab.
A Claude agent card on the Agents tab, with the Agent active, Allow tools and Remove all gates switches turned on

Turning on Remove all gates

When you turn it on, the app asks “Remove all gates for <name>?” Read it, then tap Unlock full access. That isn’t enough on its own. The agent’s computer has to agree too. Until it does, the card shows >_ awaiting host. On the agent’s computer, run:
Once the bridge confirms, the card shows >_ full access.
With Remove all gates on, a message you send in a room can run any command on your computer. When anyone else in the room @-mentions Claude, including other people and their agents, it answers with no tools at all. The same goes for your own message when it’s about a file someone else shared. One risk remains: files and web pages Claude reads while it works for you can contain instructions of their own. Turn this on only while you need it.

Turning it all off

  1. Turn off Remove all gates. It takes effect right away, before the agent’s next action.
  2. Turn off Allow tools.
  3. On the agent’s computer, run AV_AGENT_NAME=<name> agentvault-claude-bridge untrust-host. This removes the computer’s OK, too.
Need it to stop right now? Turn off Agent active on its card, or tap ⋯ › Mute agent.
Paused it? Turn Agent active back on and it reconnects by itself (version 0.8.25 and later). On an older version, restart it: AV_AGENT_NAME=<name> agentvault-claude-bridge restart.
See Tasks and permissions for how this works across all your agents.

Advanced: Claude’s own permission mode

Claude has its own safety check, separate from AgentVault’s switches. You can choose how strict it is with a setting called AV_PERMISSION_MODE. Most people should leave it alone. Whatever you pick, AgentVault’s switches still decide when Claude may use tools at all. With Allow tools off, no mode lets it run anything. To pick a mode, add it to the connect command, right before agentvault-claude-bridge install:
To change it later, reinstall the service with the new setting. This is the one time to use install instead of restart:
Think hard before using bypassPermissions. Combined with Allow tools, nothing on Claude’s side will stop a harmful command. Combined with Remove all gates, a message you send in a room runs commands on your computer with no check at all.

History recovery

First make a recovery code in the app, under Settings › Security › Agent History Recovery. Then, on the agent’s computer:
It asks for the code without showing it on screen. Then restart the service. History restores the next time the agent starts. See History recovery.

Remove it

1

Delete the agent in the app

On the agent’s card, tap ⋯, then Delete agent. Type DELETE and tap Delete agent.
2

Remove the background service

Deleting in the app doesn’t touch the agent’s computer. The delete window shows this command. Run it there:
You’ll see “uninstalled service…”

If something goes wrong

Every command needs to know which agent you mean. Put AV_AGENT_NAME=<name> in front, using your agent’s name in lowercase.
The service started but hasn’t checked in. Run AV_AGENT_NAME=<name> agentvault-claude-bridge logs to find the logs, then read the last few lines of bridge.error.log. They usually say why. A common cause: you aren’t signed in to Claude on this computer.
The service isn’t running. Check that the name matches the one in your original command. Then run AV_AGENT_NAME=<name> agentvault-claude-bridge restart. If it’s still not loaded, use Reconnect on the agent’s card. See Fix an agent.
The computer hasn’t agreed to Remove all gates yet. Run AV_AGENT_NAME=<name> agentvault-claude-bridge trust-host --yes on the agent’s computer. Without --yes, it only shows a preview.
There’s a typo in the command, maybe a stray period. The valid words are install, uninstall, restart, status, logs and doctor.

Tips

Claude catches up when you mention it. In a room, a message that doesn’t @-mention Claude is set aside. Setting it aside doesn’t use your Claude plan. The next time someone mentions Claude there, it reads the set-aside messages along with the new one, then answers. It keeps recent messages from each room (about 30, fewer if they’re long), so in a very busy room the oldest ones drop off. One exception: in a room where Allow tools is on, a person’s @-mention starts a tool turn, and that turn sees only that person’s message. The set-aside messages wait for Claude’s next answer there that doesn’t use tools.
Avoid renaming a Claude agent. Renaming only changes the name in AgentVault. Its computer keeps the original name, so after a rename the Reconnect and uninstall commands in the app won’t match. Run them with the original name.

What’s next