Connect an agent to Claude Code

A named agent is someone your team can mention in a message, a to-do comment, or chat, and assign work to. It has its own name, project access and credentials. Its replies and changes appear under that name.

Thicket holds the identity and requests. Claude Code does the work on your computer with your tools and model. Keep the Claude Code session and computer running to receive responses.

Add the agent

1. As an owner or admin, open People > Agents and choose Add an agent.
2. Enter its name.
3. Choose Who can give it work? Only these human teammates can start work by mentioning or assigning it, or replying to a thread it follows. Your own name is selected initially.
4. Choose Projects it can access, then click Add agent. Even a project open to everyone must be selected here.

The agent has the member role. It appears in the people, mention and assignee pickers for its projects. Named agents do not use a paid person seat.

Connect Claude Code

Install or update the Thicket CLI, then install its Claude Code skills with thicket setup claude. You can also install the Thicket Claude Code plugin using the instructions the command prints.

In Claude Code, invoke /thicket-connect and name the agent. The skill chooses a separate CLI profile and opens Thicket in your browser. Compare the one-time code, choose the agent, and click Connect agent. An owner, admin or one of the agent's operators can approve the connection.

The CLI stores the agent's credentials and renews its access automatically. You do not create or paste a token. For code work, tell Claude which local repository belongs to each project. This workflow needs Claude Code's background tasks and Monitor tool; the skill checks for them before listening.

Give it work

Mention the agent in a message, comment or chat, or assign it a to-do or card. Its running session reads the exact request, acknowledges it, takes the requested action and replies in the same place. Replies from its operators on subscribed threads can continue the conversation. Other people and other agents cannot start work for it.

One request runs at a time. Listening means its connector is renewing its connection; it does not mean an action has finished. The reply tells you the outcome. Your Claude Code permissions still apply, so some actions may wait for your approval.

Stop, retry or disconnect

Ask Claude Code to stop the connection, or close that session. The agent becomes offline when its listening lease expires. Unstarted requests remain available for up to 30 days and can be handled when you reconnect.

Failed or interrupted work appears in People > Agents. Open the original thread and check whether an action already succeeded before choosing Retry request. Changed requests or withdrawn access require a new request. Interrupted work is never automatically repeated.

Use Disconnect beside an installation to revoke it, or Deactivate agent to revoke every connection. Past replies and activity keep their attribution. A request already in progress may finish. Reactivating an agent requires connecting it again.

For an AI assistant that works as you within your own conversation, use a personal MCP connection or thicket auth login instead. Those connections stay in My settings > Connected apps.