Connected accounts & secrets
A runbook run needs credentials: one for the model provider, often a few for the tools the agent calls, and sometimes your own subscription so the agent runs as you. Jetty keeps these in three distinct layers, and which layer a secret belongs in determines how it is scoped and when it is read.
1. AI provider keys
Settings → AI providers (/settings/ai-providers). These are the keys that let a run talk to a model provider. Jetty supports per-provider keys for:
- Anthropic
- OpenAI
- AWS Bedrock
- OpenRouter
- Google Gemini
- vLLM
The runtime needs a key for whichever provider it's configured to use. If the runtime is pointed at OpenRouter, it needs the OpenRouter key; if it's on Anthropic direct, it needs the Anthropic one. Provider keys are scoped to the collection, and runtimes read them at run time, so you set the key once and every run in the collection picks it up.
2. Environment variables
Settings → environment variables (/settings/environment-variables). Org-wide key/value secrets and config, injected into every run in the collection. This is where everything that isn't a model-provider key lives: API tokens for the tools your agent calls, service URLs, feature flags, default config.
If your runbook scrapes with a third-party service, queries a database, or posts to Slack, the credential for that goes here. It lands in the sandbox environment the same way a local .env would, so the agent reads it however that tool expects.
3. Connected accounts
Settings → connected accounts (/settings/connected-accounts). This is OAuth credential forwarding. Instead of pasting a raw API key, you connect a subscription or account, and Jetty forwards that credential into the sandbox so the agent runs under your subscription rather than against a metered key.
Use this when you want runs to draw on a plan you already pay for rather than a separate pay-per-token key, or when the provider only offers subscription auth. The mechanics are OAuth, not a stored secret you manage by hand.
Across all three layers, secrets are never written to logs. Forwarded credentials and injected variables are available to the run but kept out of trajectory output and run logs.
Which layer do I use?
- Talking to a model? AI provider keys.
- Talking to any other service (DB, scraper, Slack, your own API)? Environment variables.
- Want the agent to run under a subscription you already hold? Connected accounts.
New here? Start with the quickstart. To see how a runtime picks up the provider key at run time, see runbook runtimes →