Connect an outside assistant

Connect an outside assistant to boxes.dev so it can find your earlier work, start Codex or Claude tasks, and continue conversations in projects you choose. The connection can access your own threads, including those started in the boxes.dev apps. It does not include teammates' shared threads or Team Template administration.

Connect your assistant

Add the boxes.dev plugin or remote MCP connection in your assistant, then:

1. Follow its authorization link and sign in to boxes.dev.
2. Check the requesting domain, return address and your account. A client name marked unverified comes from the client itself.
3. Review its requested permissions and select the projects it may access.
4. Choose Allow access to return to your assistant.

Read access includes your prompts, replies and saved conversation history. If you allow task control, the assistant will be able to start agents, send follow-ups and interrupt work using the agents' configured permissions and normal usage. Event access will let it receive notifications when work finishes, fails or needs input.

Projects you create later will not be added automatically. Reconnect if you want to change the selected projects or permissions; renewing sign-in alone will not expand them.

Find or start a task

For example, ask your assistant to find the conversation where you worked on a particular feature, read its context and suggest the next step. Search covers indexed prompts and final replies and can take time to include recent messages. Your assistant can also list recent threads when a search does not find them.

New tasks start in a new devbox using your project's saved environment. If setup or a saved environment is missing, finish that in boxes.dev first. Continuing a thread will preserve its context and may wake a sleeping devbox. Unarchive archived threads or restore stored threads in boxes.dev before continuing them. A conversation retained only as history cannot continue.

During ongoing work, follow-ups will queue by default. For Codex, ask explicitly to steer if you want to redirect its current work. Claude steering is unavailable through this connection, and its follow-ups will not appear in the app's pending-message list. If a Claude message is explicitly reported as not delivered, a new attempt needs a new request key. Your assistant can retry an identical request with the same request key after a lost response without starting a second task or sending a second message.

Follow the conversation and give input

Keep the conversation link to pick up the same work from another assistant chat. Delivery receipts let the assistant check whether a start or message was handed off. An accepted or delivered message may still be waiting to run.

The assistant can read recent saved messages and current conversation activity, then refresh as work continues. History previews may shorten large messages or omit earlier content; listed messages can be read in full. Open the thread link for very large messages omitted from the preview. Reverting or rewinding a conversation can remove messages from its current history.

Replies and activity belong to the conversation. They are not automatically matched to each request: you, other assistants and background work may all contribute. An idle conversation or a completion notice does not prove that every queued message ran.

If your assistant supports notifications, it can subscribe before starting work and read the conversation again when notified. Subscriptions need renewal within 24 hours and do not replay old events. Saved messages may arrive shortly after a notification, and background work can add later replies. Direct status checks remain available if notifications are missed.

For native approvals or interactive questions, open the thread link and respond in boxes.dev. The assistant connection does not answer those dialogs. Some questions cannot be interrupted remotely; follow the thread link when the assistant reports that limitation.

Stop current work or disconnect

Ask the assistant to stop current work or a pending startup in the conversation. Interruption will preserve files and queued follow-ups, and the next queued message may start immediately. It will not withdraw a message already in the agent's queue or stop all future work. A Codex prompt interrupted before it appears in history may return to your boxes.dev composer as unsent text.

On older devboxes, Codex's first response can appear idle while it is still running until the devbox receives the startup fix. Claude can also reject a stop before its first activity is visible. If the assistant cannot confirm a stop, open the conversation to inspect its work; an idle status alone does not confirm it finished or stopped.

To remove the connection, open Account → Assistant connections, choose Remove access, and confirm. New requests and future event deliveries will stop, but accepted tasks will keep running. Interrupt those tasks separately if you also want them to stop.