Workspace settings
Choose the agent's provider and key, set its monthly budget, decide about knowledge, and set per-project autonomy.
A workspace admin runs the agent from Settings, AI. Anyone with ai:read can open that page and see the workspace's AI budget. You need ai:manage to change anything there, and to see the agent's panel at all. Everyone with agent:use can see the agent's requests and talk to it.
Provider and key
Pick one of four.
| Choice | Runs on |
|---|---|
| Server default | The provider, model and key the operator set under ai. The page shows which |
| Anthropic | Your own Anthropic key |
| OpenAI | Your own OpenAI key |
| OpenAI-compatible | Any endpoint that speaks the OpenAI API. You add a base URL too |
Paste the key into API key and choose Save provider. After that the page shows only Key set, ends in wxyz. Nobody can read it back, including you. Replace swaps it for a new one, and Remove deletes it and sends the agent back to the server default.
Keys live in Vault, never in the database. A key the agent uses has to be available to every project, because the agent works across the workspace, so you can't narrow it to some. You can't delete it while the agent runs on it either.
The base URL for an OpenAI-compatible endpoint is checked when you save. It has to be an absolute URL with no credentials in it, and in production it has to be https. Loomplane refuses addresses on private networks and loopback outside development, link-local and cloud metadata addresses everywhere, and anything in planning.endpoint_denied_cidrs, because the agent sends its key there. The refusal names the rule.
The Model field starts from the model the agent runs on now. Type another to change it. Loomplane refuses a model it has no price for and says so under the field.
Budget
There are two budgets, and the order matters. The workspace has a monthly AI budget that everything AI shares: models, embeddings and the agent. The agent gets a share of it. Neither is set when a workspace is new, and the agent can't be turned on until both are.
-
The workspace's budget. It's the AI budget section at the top of this page. You'll see the budget, how much of it is shared out, what's left and who holds each share. Type an amount into Workspace budget (USD) and choose Save. It can't go below what's already shared out. The API is
PUT /api/v1/ai/settings/budgetwithmonthly_budget_usdand therevisionyou read fromGET /api/v1/ai/settings. A stalerevisionis refused, so read again and retry. -
The agent's budget. Further down, in the agent's panel, give the agent a monthly budget in USD and choose Save budget. It comes out of what's left of the workspace's budget, and the field tells you how much that is. Ask for more and the page stops you before saving and tells you to raise the workspace's budget first. The page shows what the agent has spent this month beside the limit. The API is
PUT /api/v1/ai/settings/agent/budgetwithmonthly_budget_usdandrevision. -
The switch, below.
You need both budgets before you can turn the agent on. When the month's spend reaches the limit, the agent's model calls are refused with a message that says the workspace has used its AI budget, and they stay refused until next month or until an admin raises the limit. Work that's already waiting on a person isn't lost. Setting the budget to 0 releases it, which you can only do while the agent is off.
Knowledge
The Knowledge switch lets the agent search documents before it answers. When the server can't support it, the switch is disabled and the reason sits beside it. Knowledge lists each reason and what fixes it.
Turning it on and off
The switch at the top of the panel turns the agent on for this workspace. If something stops it, the message appears under the field it's about: a missing budget under the budget, an unpriced model under the model. Turning it off cancels running work and pending requests, so don't do it in the middle of a release you care about.
Autonomy, per project
Open a project's Settings tab. When the agent is on, an Autonomy card sits beside Responsible team.
Autonomous promotion says whether the agent may promote a release to production without asking first, once someone has asked it to release and every gate is green. It starts off for every project, so new projects ask first. The minimum staging soak (none, 30 minutes, 1, 2, 6, 12 or 24 hours) is how long staging has to have run the same commits, healthy, before the agent may promote them.
With the switch on, the agent promotes only a release someone asked it to prepare, and only when all seven gates hold. Release preparation lists the gates and what always goes to a person. With the switch off, the agent still asks before production.
Both need deployment:approve on the project. Without it you see the card with the controls disabled. The agent can't change either setting, whatever it's asked. The same call is PUT /api/v1/projects/{id}/autonomy with autonomous_promotion and min_staging_soak_seconds, which must be a multiple of 60 up to a week.
Deleting or replacing anything in production that holds data always waits for a person, and the switch never changes that.
Requests
Anything the agent asks of a person lands on Deployments, Agent. Who sees what is on What needs a person.