Skip to main content
Three separate choices determine how AI work runs:
  • Agent: the software doing the work—ReasonOS (built-in), Claude Code, Codex or Gemini CLI.
  • AI model: the model that software uses to reason.
  • Account or key: the credentials that provide access to that model.
Adding an Anthropic API key does not install or select Claude Code. It makes compatible models available to the built-in ReasonOS agent. You choose agent software and a model separately.

​
Connect your own account

Open Settings → Your account → Profile & preferences, then Personal AI accounts. This section is available with the core ReasonOS product. Each provider offers its supported sign-in or token-paste flow. Follow the instructions on its card; never paste credentials into an issue or chat.
  • Claude: a connected account can serve Claude Code and the built-in ReasonOS agent.
  • OpenAI: a ChatGPT sign-in serves Codex, not the built-in agent. A personal OpenAI API key can also serve the built-in agent. A ChatGPT subscription and API access are different credentials, not interchangeable model choices.
  • Google: a connected account can serve Gemini CLI and the built-in agent.
The link belongs to your ReasonOS account, not to one project or branch. Managed agents use your identity to select your credentials. Usage remains subject to the provider’s access rules and limits; “Your AI account” does not mean every request is free or covered by a subscription.

​
Configure a shared project key

Open Project settings → AI provider keys. With permission to manage project secrets, add the provider’s API key there. Other people working on that project can use it through the built-in agent when they have no compatible personal credential. Project keys also support project AI workflows. For ordinary setup, use the shared provider key. Additional named keys are optional advanced configuration for workflows that need to select particular credentials. Giving a key a name does not make it a different model or grant extra capabilities. For the built-in agent, the automatic credential order for a provider is:
  1. Your compatible personal credential.
  2. The project’s shared provider key.
  3. A single eligible named key for that provider in the project, then in the organization. A scope with multiple eligible keys is skipped rather than choosing one arbitrarily.
  4. The deployment’s Anthropic credential, if configured; this last fallback does not apply to every provider.
The built-in model picker shows the resolved source once as Account or key. CLI agents use their separate vendor account configuration and do not inherit the built-in agent’s project-key fallback. If a CLI needs account access, connect the corresponding personal account rather than adding more project keys. In managed mode, vendor credentials are held by ReasonOS’s credential service and model calls are relayed; the original vendor key is not installed on the branch server. The server uses a scoped session credential. This is not a promise of isolation between untrusted people sharing a branch: collaborators still share its execution environment. Only invite people you trust to it.

​
Where defaults come from

These settings have different scopes:
  • Organization Settings → Agent installation: which vendor CLIs branch servers install and their versions.
  • Project settings → Default agent: the starting agent for new branch servers. Already-running servers retain their current settings.
  • Branch AI settings: the branch’s shared defaults. Changing the default agent does not replace the agent in existing conversations. Model and tuning changes in this shared settings page can affect running sessions; use a conversation’s own model picker for a conversation-specific change.
  • AI Work setup or a conversation’s model picker: an optional agent/model choice. The browser can remember a choice for future conversations, but it is not an account-wide default synced to other browsers.
In AI Work, Use branch defaults clears only the browser’s remembered model choice for new conversations. Existing conversations keep their own agent and model. It does not disconnect accounts or change project keys.

​
If a choice is unavailable

  • Check whether the agent is installed on the branch server; an API key alone does not install a CLI.
  • Check whether the connected account can serve that agent and model. A ChatGPT sign-in does not provide native OpenAI API access.
  • If the pane reports an incompatible remembered choice, select a compatible model or choose Use branch defaults.
  • If a deployment has not enabled the required model relay or provider flow, ask your administrator. Saving another key does not fix missing deployment configuration.
Standalone, self-run servers use their own local credential configuration; these managed-account instructions do not replace it. See Using the agent for command-line use.