Assistant
The Rivets assistant is a conversational front end to Rivets. You tell it what you need, and it gets the work done by delegating: it creates tasks, watches them, reads what they did, and reports back. It is not a coding agent. The coding agents are the tasks it starts.
Where it lives
| Client | How to open it |
|---|---|
| Web and Mac | Press ⌘J, or choose the assistant in the sidebar. It opens as a panel in the bottom-right corner that you can minimize or maximize. |
| iPhone | The Assistant tab. |
| iPad | A sheet. |
| Slack | Mention @Rivets in a channel or message the app. See Slack. |
With no conversation open, the assistant suggests a few starting points: Create a task, What’s running right now?, Summarise my open Linear issues, and Check on my pull requests. Choosing one fills the composer without sending.
Assistant conversations are private to you.
What it does
The assistant works through the same tools Rivets exposes over MCP:
- Creates tasks. For engineering work, it writes a clear, self-contained prompt and starts a task in the right repository, with a workspace, a branch, and a coding agent.
- Watches them. Right after starting a task, it asks Rivets to wake it when the task finishes. It does not poll. When the task completes, fails, or is cancelled, the assistant reads the output and tells you what happened and what it recommends next.
- Answers questions. It lists repositories and tasks, and reads task output before reporting, rather than guessing.
- Works with Linear. If Linear is connected, it can find issues, start one task per issue linked back to it, and comment on or update issues when you ask.
- Checks pull requests. It can read a task’s pull request state, checks, and reviews.
It does not commit, push, or open pull requests from its own workspace. Those happen in the tasks it starts.
Which model it uses
The assistant runs on your worker, with your agent login, like everything else in Rivets. Rivets bills no inference.
It uses your default agent from Settings › Agent defaults, but always on that agent’s cheapest model at the lowest reasoning effort: Haiku for Claude Code, Luna for Codex. The assistant only reads short messages and calls tools. Real work happens in the tasks it starts, which use your chosen model.
Requirements
The assistant is available when:
- The organization has at least one connected repository.
- A worker is online that can run the assistant’s tools.
- That worker has a signed-in agent ready.
If any of these is missing, the panel says which one. If the assistant’s worker goes offline, Rivets moves the conversation to another available worker and continues.
Start over
Start a new conversation from the panel header. The header’s menu also lets you refresh, archive the conversation, or reset the assistant’s workspace.
Customize its prompt
The assistant’s system prompt is the Rivets assistant prompt under Settings › Prompts. Edit it for your account, or as an owner for the whole organization. It supports {{repoName}} and {{workspacePath}}. See Custom prompts.