/ Get started

Connect your AI models

The AgentWork team thinks with AI models you own, through API keys from a dozen providers, coding plans or subscriptions such as Claude and ChatGPT, so model costs go straight to your account.

AgentWork does not resell model access. You connect your own subscription or API keys, and model usage is billed to you directly by the provider. Connections live in Settings → Models & Usage → Connections.

Connection options

Providers are shown as cards, in four sections by how you pay for them. Click Connect on a card: a window shows where to get the key, step by step, with links to the provider's console. The endpoint fills itself in, so you only paste the key.

Kimi appears three times on purpose: signing in with a Kimi Code membership, the API key of that membership, and a pay-as-you-go key from the Moonshot platform are three different things.

Subscriptions and coding plans use your plan's quota. Some providers limit how their plans are used in other apps, so your account may be limited by the provider. Use them at your own risk. Pay-as-you-go API keys have no such limits.

Signing in with Gemini goes through Google's sign-in for Antigravity. After you approve, the tab goes to a localhost page that does not load: that is expected. Copy that page's whole address and paste it in AgentWork. Cursor only needs you to approve on its page; AgentWork connects by itself.

Signing in with a subscription

Click Connect on the provider's card under Subscriptions, then Sign in with ….

• ChatGPT, Kimi Code, GitHub Copilot and Kiro show a link and a short code. Open the link, sign in, and enter the code there. Leave the AgentWork window open: it connects by itself once you approve, with no need to click anything.
• Claude opens a Claude page. Sign in and approve; the page then shows a code. Copy it into the Code field in AgentWork and click Connect.

The account's tokens are kept encrypted on your instance and renew themselves; AgentWork never sees your password. Each connection can use that account's models only. For ChatGPT, AgentWork lists only the models your plan offers; if you change plans, remove the connection and sign in again.

If ChatGPT says device sign-in is turned off, turn on device code sign-in in ChatGPT's security settings and try again.

Claude subscription

Signing in (above) gives every agent your Claude models. Agents that run on Claude Code use a token instead:

1. On your computer, install Claude Code and run claude setup-token.
2. Sign in with your Claude account. The command prints a token.
3. In AgentWork, click Connect on the Claude card under Subscriptions and paste the token below Or paste a token instead.

The token is stored encrypted in your instance and is never shown in full again. The team's usage counts against your subscription's quota. When the quota runs out, the agents affected move to the repair bay in the Office and carry on once the quota is back.

API keys

Create a key on the provider's site and paste it in the window that opens from the provider's card. AgentWork checks the key and lists the models it can use. We recommend a key just for AgentWork, so its usage is easy to follow and it can be revoked without affecting anything else.

Your own endpoint

If you run a model gateway of your own that speaks the OpenAI API, or use a provider's regional endpoint (such as Moonshot's or Qwen's China endpoints), choose Your own endpoint under Other and enter its address. It must be reachable from the internet over HTTPS. Addresses on a local network or your own computer are refused for security.

One provider or two?

One connection is enough to start. Two providers from different model families, for example a Claude subscription plus an OpenAI API key, give the best results: QA can then review the work with a different model from the one that did it, which catches mistakes typical of one model. See QA review.

Choosing models per role

In Models & Usage → Routing, you choose the model for each role: specialists, reviewer, second reviewer and more, including fallback models for when the main one is unavailable. After onboarding, every role uses the model you picked in the connection step.

Checking which models answer

A provider or router sometimes lists models your account cannot use. On a connection's card, Check models sends each model one short test message (it uses a little of the connection's quota). Models that refuse it for good, for example not served to your account or not found, are hidden from every list of models, with the provider's reason under Which and why. Models that are only busy or limited for now stay listed. Show hidden models again lists them all once more; a later check may hide them again.

When an answer fails, the chat says why, as the provider put it (for example "the model was not found" or "verify your account"), and where to choose another model for that agent.

Costs and limits

• Model costs are billed to your account by the provider, at their prices and terms.
• A budget per project. Every project has a token budget: by default 500 thousand (Quick), 1.5 million (Standard) and 4 million (Deep). Change these in Models & Usage → Budgets.
• An 80% warning. When a project reaches 80% of its budget, a warning is recorded. When the budget runs out, the project stops and its lead tells you in the chat. Retry incomplete tasks grants one more budget and the tasks resume where they stopped.
• Follow usage in Models & Usage: model calls, tokens per agent and per project, and the work that uses the most tokens.