Skip to main content

The Org Agent

The Org Agent button in the panel's sidebar appears once your organization's agent is provisioned (organization admins provision it under Settings → AI → Org Agent; on some plans it is provisioned for you).

The agent runs in its own isolated environment per organization and holds two narrowly scoped credentials: one that lets it call models (through the same plan-capped lane as Platform AI), and one that lets it read your organization's projects, applications, and deployments through the platform API. It cannot modify anything, cannot see other organizations, and cannot reach the internet.

On paid Claude models the agent's system prompt is prompt-cached by the gateway (see Prompt caching), so the repeated prefix of a multi-step tool-call turn is billed at the cache-read rate rather than full price.

Conversations are request/response rather than streamed — on the free tier the models run on CPU, so a reply can take up to a minute; the panel says so while it works.

Running the agent on your own key. By default the agent uses the platform's models on your plan. An organization admin can instead run it on one of your organization's own AI providers (the ones added under Settings → AI): pick it as the agent's model under Settings → AI → Org Agent and provision. The agent keeps the same locked-down tools and isolation; only the model changes. Usage on your own key is billed by your provider and does not count against your plan's AI spend cap.

Providers with an OpenAI-compatible API are supported: OpenAI, Anthropic, Mistral, DeepInfra, Perplexity, and custom OpenAI-compatible endpoints. If you change that provider's key, provision the agent again to pick it up.

Step by step: run the agent on your own provider​

An organization owner or admin does this once:

  1. Open Settings → AI and click Add AI. Fill in Name, API URL (your provider's OpenAI-compatible base URL, for example https://api.openai.com/v1 or https://api.anthropic.com/v1), API Key and Model, and save.
  2. On the same page, in the Org Agent card, open Agent model and pick that provider. Providers marked (not OpenAI-compatible) can't be chosen.
  3. Click Provision, or Re-provision if the agent already exists.
  4. Check it worked: the card's Status reads running and the model reads your provider. Then open the /ai panel, choose Org Agent, and ask "what's deployed right now?".

Only owners and admins can add, edit or delete providers; members pick from them when starting a chat. If you change a saved provider's API URL, enter its API Key again before loading its models — the stored key is only ever sent to the URL it was saved with.

To go back to the platform's models, pick Platform AI (your plan) and click Re-provision. Tear down removes the agent and revokes its credentials; your providers are kept.