Connecting a provider
Connect one provider on the host that will run your session. You need an installed, running hub and either a hosted account with suitable access or a reachable local model server. No existing API key is required for the Ollama path.
The first-task walkthrough uses this dialog without leaving New session. Settings → Providers & credentials manages the same host connections and credentials. If access is already configured, choose that connection rather than adding it again.
Start your first session
- Open New session. Choose a working directory and, optionally, write your task. These stay in the composer while you connect a provider.
- Select Connect provider, then choose a provider. Show all providers reveals the complete catalogue and its search; local and custom endpoints have their own routes.
- For a standard API connection, paste the actual API key, not an environment-variable name. The field is masked. Use Get an API key beside the form, or these provider pages: Anthropic, OpenAI, Gemini, OpenRouter.
- Select Save and check. Evener saves access on the Evener host, reloads the resolved connection, and requests the provider's model list only. It does not send your prompt or make a generation request. A returned list does not prove generation access, quota, billing status, or a model's tool capability.
- After a successful check, select Choose model. The model chooser opens with models returned for that connection's actual instance name. Choose a model; Show all models returns to the full catalogue. If no models are listed, check the provider's model access and endpoint rather than guessing a model ID.
- Review your working directory, task, and explicit model. Review Effective access near Start and use Change if the resolved permission is not what you intended. Instructions in the task do not restrict tools; the access setting does. Then press Start. Connecting and choosing a model do not start a session or change the global provider default.
The task and working-directory drafts stay in New session while the provider dialog and model chooser open, fail, retry, and close. Check them again before Start. A successful credential save survives a later model-list failure; retry the check without pasting the key again. Cancel does not start a session.
Subscriptions are not API keys. Claude, ChatGPT, and Gemini subscriptions do not automatically include separately billed API access. OpenAI API keys and OpenAI Codex sign-in are separate connections. For the latter, choose its sign-in route and follow the device or browser/redirect instructions instead of pasting an API key. Availability depends on your account and provider terms.
Other access routes
- Cloud providers: use Show all providers, choose Azure, Vertex, Bedrock, or another provider, then Configure provider when offered. Supply the required project/resource/region or endpoint variables. The full editor retains protocol, surface, header, and custom-name controls. Vertex can use credential JSON or existing Application Default Credentials; follow the selected provider's access instructions rather than substituting an API key.
- Local/custom endpoints: start the server on a host reachable from the Evener host and use the local/custom route. Configure its actual URL and any required credentials. An empty optional key means keep resolved host access, not disable authentication. See Ollama for local discovery.
- Already configured on this host: expand Already configured access on this host? and choose Manage existing connections. The connect dialog closes and Settings → Providers & credentials opens; each instance row there opens the complete inspector, where you can review the displayed destination and credential source and test current access before contacting the provider. The guided form also offers Use existing host access under Advanced settings.
- Advanced management: Settings → Providers & credentials → Connect provider uses the same connector. After a successful check, Done returns to Settings; it does not open the New session model chooser. The instance rows are the real editor, with custom instances, rename, reset overrides, replace/clear stored credentials, logout, source inspection, and explicit default management. In the session model chooser, only an explicit model pick requests a session model switch.
Connect from the CLI
For ChatGPT/Codex access, run these on the execution host:
evener openai login
evener openai statusComplete the browser or device flow privately. These commands use the openai-codex instance by default, not the API-key openai instance. On a headless host, EVENER_LOGIN_HEADLESS=1 evener openai login requests the device flow. Status checks local login state; it does not prove quota or model generation.
For API-key access, the browser's masked form is the simplest way to store a key without shell history. A service can instead receive the provider's supported variable from its secret manager. Restart the service after changing that environment. Do not type a literal secret into --credential-header or another command argument.
Inspect what resolves without making a generation request:
EVENER_OFFLINE=1 evener providers list
EVENER_OFFLINE=1 evener models inspect openai/gpt-5.6Substitute the instance/model you intend to use. For a custom gateway, author an instance without probing it:
evener providers add work --base openai \
--base-url https://gateway.example.com/v1 \
--api-key-env WORK_API_KEY --no-probeThis writes configuration; it does not set WORK_API_KEY or validate that example endpoint. Supply the real gateway URL and inject its key securely. An instance pointing to a different endpoint does not inherit vendor key variables. Inspect the resulting destination before any live check. Without --no-probe, providers add may send provider requests.
Once the connection resolves, use an explicit instance/model in the CLI, or choose it in New session. Review the result using Daily work.
Repair or return
Save failures keep the current credential draft available for retry. Check failures do not undo a successful save: inspect the inline error, confirm the destination and active source, repair access or endpoint configuration, and retry. Changed destinations require review rather than silently reusing inherited access. If a provider cannot check model-list access, Continue without verification is an explicit unverified route, not a success badge.
Change provider returns to discovery; it does not carry your secret to a different provider. Cancel returns to the originating screen without choosing a model or starting a session. Your composer draft remains, but credentials or configuration already saved on the host are not rolled back by Cancel.
For storage, environment precedence, and launch internals, see the technical provider reference.