Integrations: your own model provider, alerts that open issues, tickets agents read
A new integrations service holds a workspace's connections to outside systems, with their secrets sealed under AES-256-GCM and never returned. Models: a workspace can send its agents' model traffic to its own Anthropic key or any Anthropic-compatible endpoint. Every sandbox now reaches models through a new model proxy (models.g1t.sh) with a token for its one run; the proxy adds the credentials, so no sandbox holds a key, g1t's or the workspace's. Runs on the workspace's own provider are charged a flat orchestration fee ($0.10) instead of model cost plus margin. Alerts: Sentry (signed webhooks, the latest event's stack trace, resolved when the fix merges), Datadog and signed generic webhooks open one issue per problem however often it fires, reopen it on a regression, and can put an agent on it at once. Trackers: Jira and Linear keys such as TECH-1234 resolve to the ticket. Agents get referenced tickets in their starting context and a get_context tool; people import a ticket as a linked issue; tickets hear back when work starts and lands. API: list, connect, test and disconnect integrations, get_context, import_issue, and POST /hooks/{integration} for outside systems.
- Also:
- the API reference's sidebar no longer slides under the header.
No changes