v0.2.0

Released 2026-10-09.

Alerts and notificationsCopy Anchor Link

• Rules for budget thresholds, spend spikes, error rates and failing connections, for the installation and for each Team and Project, checked by a background evaluator.
• Built-in budget alerts for personal workspaces, for their owner only.
• In-app notifications with a bell in the top bar, and email through the SMTP relay. Unknown cost is flagged, never counted as zero.

See Alerts.

Keys and modelsCopy Anchor Link

• Key safety: findings for keys with no expiry, no budget, a holder who left, too long a lifetime, an old secret, no use, or access to every model, each with a fix. Personal keys appear in Admin only as counts.
• Model compare: two to four models side by side, with exact prices, ceilings, readiness and 30 days of activity.

IdentityCopy Anchor Link

• SCIM 2.0 provisioning for users and groups, with deactivation that suspends and revokes keys, and group grants from pushed groups.
• Signing-key refresh: the identity provider's keys are cached within bounds, refetched once on an unknown key, and kept for a bounded time if the provider is unreachable.

OperationsCopy Anchor Link

• Prometheus metrics on a separate listener, with low-cardinality labels, plus example alert rules and a Grafana dashboard.
• Readiness checks for the database, schema and dashboard.
• scripts/backup.py: checksummed backups and guarded restores.
• A load test against a mock provider with exact ledger checks. It found and fixed connection-pool exhaustion at 200 concurrent requests (GATEWAY_DATABASE_MAX_CONNECTIONS).
• Budget totals: budget checks read maintained per-period totals instead of scanning history. In the load test with 200,000 requests in the month, throughput went from 19 to 124 requests per second. budget verify checks the totals.
• A browser test suite for every persona, with axe accessibility checks, in CI.

New workloadsCopy Anchor Link

• Realtime audio: /v1/realtime, a WebSocket proxy for OpenAI's Realtime interface, with an event allowlist, per-response budgets and session limits.
• Async jobs: video generation and batches of Chat Completions on OpenAI, polled in the background.

LogsCopy Anchor Link

• The model the provider says it served is recorded beside the configured one, and Anthropic's thinking tokens are counted as reasoning tokens.

UpgradingCopy Anchor Link

Migrations 0011_alerts, 0013_upstream_model, 0014_scim, 0015_budget_totals, 0016_async_jobs and 0017_realtime. See Upgrades.