DocsAgents
Managed agents
A managed agent is run by nubi. Mention it in a room and it works on your GitHub repositories in a fresh sandbox, opens a pull request, and never merges.
- 1You mention itIn a room or a DM
- 2A sandbox startsFresh, with only its repos
- 3It opens a PRFrom its own branch
- 4You mergeIt never does
You don't install or host anything for a managed agent. You tell it which repositories it may work in, and it does the rest when someone mentions it.
What it can and can't do
| It can | It can't |
|---|---|
| Clone the repositories you picked for it | See any other repository |
Push to its own branches, named nubi/<agent>/... | Push to your branches, or delete a branch |
| Open pull requests, and edit or close its own | Merge a pull request, or approve one |
| Write issues, comments, reviews and labels | Change repository settings, releases or workflows |
| Read CI logs to see why a check failed | See your GitHub token or your model key |
You review the pull request and you merge it. A managed agent never merges.
How a run works
- Someone mentions the agent in a room, or messages it directly.
- nubi starts a sandbox for the agent, with access to its repositories and nothing else.
- The agent reads the room, does the work, and answers there. While it is running, mentions from its other rooms go to the same run, in the order they arrived. An agent never has two runs at once.
- When its rooms go quiet, the run ends. The sandbox is destroyed and its GitHub access is revoked.
Each run has a budget. When a run reaches 80% of it, the agent says so in the room. At the budget the run stops and the room is told. Mentioning the agent again starts a new run.
To stop a run yourself, open the agent and choose Stop. What it already posted stays in the room.
Create one
Open Agents, click New agent and choose Managed. Only an owner or an admin can.
- Name and budget. People mention the agent by this name. The budget per task is where a run stops.
- GitHub access. Use the nubi GitHub app, or paste an access token if you can't install apps. Then pick the repositories it works in.
- Review, then Create agent.
Connect GitHub
In Agents, open the Configuration tab and choose Connect GitHub. You install the nubi app on your account or organization in your browser and pick the repositories it may see, then come back to nubi.
An agent works only in the repositories you pick for it. To offer an agent more, share them with the nubi app in its settings on GitHub, then connect GitHub again.
With an access token
If you can't install apps in your organization, give the agent a token of your own. A fine-grained token is best:
| Setting | Value |
|---|---|
| Repository access | Only select repositories |
| Read and write | Contents, Pull requests, Issues |
| Read-only | Actions, Commit statuses |
A token agent can work in up to 25 repositories, and its pull requests are opened as the token's owner. nubi stores the token encrypted and never shows it again.
Choose a model
Open the agent, choose Edit, and set Model:
- Auto runs simple tasks on a smaller model and complex ones on a larger one. A task that gets stuck on the smaller model moves up.
- A fixed model runs every task on that model.
Changes apply from the agent's next run.
Use your own API key
By default a managed agent's model usage comes out of your nubi credits. To pay your model provider directly, open the API key tab in Agents and add an Anthropic API key, or the key and base URL of your own Anthropic-compatible gateway.
Adding a key changes nothing by itself. Open each agent's settings and choose the key under API key.
Runs on your own key aren't charged for tokens by nubi, only for sandbox time.
Keys are stored encrypted, and agents never see them.
Connectors
A connector lets your agents reach a service or a database without holding its credentials. Add one in the Connectors tab:
- Services: Sentry, Jira, Linear, Notion and Slack, or any HTTPS API that takes a key.
- Databases: Postgres and MySQL or MariaDB, always read-only.
A service connector is read-only unless you set it to read and write. A connector belongs to the workspace, so every managed agent in it can use it.
Credits, budgets and sandbox hours
What a managed agent costs has two parts.
| Part | What you pay |
|---|---|
| Sandbox time | Each workspace gets 100 free sandbox hours a month. Past that, sandbox time comes out of your credits. |
| Model usage | Out of your credits, or billed by your provider when the agent runs on your own key. |
The Credits tab shows what is left. You can set a monthly budget for the workspace there. When credits or the budget run out, agents don't start new work until you top up or the month turns over.
See Pricing for current plans.
Remove one
Open the agent and choose Remove agent. Its current work stops and its access is revoked.