Google Vertex AI is now available as an AI Gateway provider, and dataset items can be filtered by status. Gemini decision models are supported in decision-model evaluators, and patterns in the Topics view now include on-demand map refresh plus pinch and ctrl+scroll treemap zoom.
Langfuse
npx @buildinternet/releases get langfuseThe AI gateway now supports Microsoft Azure, Google AI Studio, and custom base URLs for OpenRouter connections, and gateway keys can be restricted to model patterns. Sessions are now bound to the user id they were issued for, and SSO-domain users are sent to SSO from invitations and sign-up.
OpenRouter, xAI, and Vercel AI Gateway are now generally available as AI Gateway providers, and custom base URLs and extra headers are available on all deployments. Audit logs record read-access events for MCP tools, monitors gain a SUBSTITUTE_OK no-data mode, and trace transcript evaluations can be scheduled.
Added a USER API key scope evaluated against user and key policies, and an external HTTP code eval dispatcher for evals. Bedrock is now available in every environment through the AI Gateway, custom models can be offered everywhere, and topics setup can filter traces.
Changing the bucket or endpoint URL now requires header values and storage secrets again, and gzip exports no longer stall when the uploader reads late.
External media storage is now an enterprise feature, and the AI gateway adds an Amazon Bedrock upstream with API key auth plus custom base URLs and extra headers for OpenAI and Anthropic connections. Amazon S3 buckets and exports now support IAM role auth.
The AI gateway is now included in the self-hosted Docker Compose setup, and Topics summaries, assignments and rules are available in the public API, MCP and in-app agent. Audit logs now record read-access and export events, plus dozens of smaller fixes across prompts, exports and sessions.
AI Gateway rate limits for Pro, Team, and Enterprise plans are raised to 10,000 requests per minute and 5 GiB per minute. Ingestion now returns 503 with Retry-After when event processing is throttled, prompt dependency resolution caps resolved prompt size, and Codex setup is simplified to an exported key and config.toml.
The legacy traces export now supports opt-in trace field groups for blob export.
AI Gateway /resolve now enforces per-organization request and byte rate limits, and the Admin API can transfer projects between organizations. Audit logs now always record support access reads and mark them in ClickHouse.
The Admin API can now transfer projects between organizations.
An AI_GATEWAY organization membership role lets teams invite gateway members and filter gateway analytics by user and API key, and skills version automations are now supported. Legacy public API reads are throttled per minute again, and storage secrets must be re-entered when the bucket or endpoint changes.
Alerts, project notifications, and prompt changes can now be sent to PagerDuty with the routing key on the automation, and evals gain trace-level filter conditions. Also fixed OTel usage mapping double counting the cache-write TTL breakdown, plus a batch of AI Gateway additions including an analytics tab and onboarding flow.
AI gateway now reports product analytics for setup, credentials, invites, keys, and model access, and records model, key owner, byte, and upstream error facts on generations. Adds a self-hosted government plan for lf_gov_ license keys, an invite-ownership check, annotation queue validation, and patched Node.js RPMs in the web and worker images.
Invites now verify that the invite's project belongs to the organization, and webhook and PostHog destination changes require secrets again. Legacy webhook headers are now masked and encrypted, and annotation queue items are validated when added.
The release is rebuilt to pull in Node.js patches with security implications. The only other change skips Azure LLM connection tests in the worker until the CI key is replaced.
Reopening native media chips now renews expired signed URLs instead of failing, and the AI gateway's telemetry flush gets its own shutdown budget. Topics view also works better on mobile, and retired legacy OpenAI snapshots were removed from the model picker.
Changing a webhook or PostHog destination now requires secrets again, and the public API gains unstable read-only Topics endpoints. Also added a runBatchEvaluation MCP tool with batch_action_id on batch run outputs, and Cmd+K jumps to prompts, datasets, queues, and evaluators.
MCP can now create skill versions and manage their labels and tags, with available and loaded skill name filters added to trace filtering. Legacy webhook headers are now masked and encrypted, and the AI gateway gains guided setup plus per-tool setup instructions after key creation.
Members can now create their own AI gateway API keys, and self-hosting gains a global enable flag with derived JWT public key and key ID. Fixed LIKE wildcard escaping in eval and dataset filters, revoked gateway keys when a user leaves the org, and restricted audit logs to enterprise plans.

