Skip to main content
​
v2.1.46

​
Changelog

  • Removed GitHub App authentication from GitHub Copilot provider keys: Bifrost now ignores github_copilot_key_config (app_id, installation_id, repository_id, private_key, github_domain). Migrate before upgrading: put a fine-grained personal access token (github_pat_…, with the Copilot Requests account permission, from an account with an active Copilot plan or assigned Copilot seat) or a GitHub OAuth token (ghu_… / gho_…) in the key’s value; a key that relied on the App fields no longer authenticates.
  • Added bifrost.mcp.maxProtocolVersion (mcp.max_protocol_version), one of auto, 2026-07-28, 2025-11-25 or 2025-06-18: the newest MCP protocol version offered to MCP servers. bifrost.mcp.clientConfigs[].maxProtocolVersion (mcp.client_configs[].max_protocol_version) overrides it for one server. Upgrading a deployment that already has MCP servers caps them at 2025-06-18, the version they used before; set auto to offer the newest version.
  • bifrost.mcp.clientConfigs[].tokenExchange (mcp.client_configs[].token_exchange) without clientId is now refused unless useIdpCredentials: true or profile: id_jag. Such a config never worked: Bifrost’s own config validation rejected what the chart rendered.
  • Added profile: id_jag to bifrost.mcp.clientConfigs[].tokenExchange (mcp.client_configs[].token_exchange.profile) for Cross App Access to MCP servers (Enterprise): Bifrost gets an ID-JAG for the user from the identity provider and redeems it at the MCP server’s authorization server. Set audience to that authorization server’s issuer and resourceClientId (resource_client_id) to Bifrost’s client there. resourceTokenEndpoint is discovered from the issuer when left empty. Add resourceClientSecret or resourcePrivateKey unless Bifrost is a public client.
  • Added bifrost.mcp.clientConfigs[].promptsToExpose and .resourcesToExpose (mcp.client_configs[].prompts_to_expose / resources_to_expose): which of an MCP server’s prompts and resources the MCP gateway serves, ["*"] for all and [] for none. A new server serves everything when the field is omitted; on an existing server, omitting it keeps the stored list.
  • Added openapi as a bifrost.mcp.clientConfigs[].connectionType, configured by openapiConfig.spec, .specFile, .specUrl, .baseUrl, .securityCredentials, .includeDeprecated and .maxResponseBytes. Renders into mcp.client_configs[].openapi_config.spec / spec_file / spec_url / base_url / security_credentials / include_deprecated / max_response_bytes. Bifrost builds an MCP server from an OpenAPI or Swagger document, with each API operation as a tool; operations the document marks deprecated are skipped unless includeDeprecated: true. specFile is read relative to bifrost.appDir, so mount it with volumes/volumeMounts. authType can only be none, headers or per_user_headers. The connection type existed in Bifrost but had no Helm path, so it could only be added from the dashboard.
  • Added storage.logsStore.objectStorage.format and .compression (logs_store.object_storage.format / compression) to write offloaded log payloads as Parquet instead of JSON. format is json or parquet; compression is none or gzip for JSON, and none, gzip, snappy or zstd for Parquet. Parquet works on S3 and Azure; GCS stays JSON-only.
  • Added azure as a type for storage.logsStore.objectStorage and bifrost.auditLogs.objectStorage (logs_store.object_storage / audit_logs.object_storage): bucket is the Azure Blob container, with either endpoint (Azure identity) or connectionString. For logs, existingSecret with connectionStringKey reads the connection string from a Kubernetes Secret.
  • Added storage.logsStore.matviewRefreshTempFileLimit (logs_store.config.matview_refresh_temp_file_limit), for example "10GB": Postgres temp_file_limit for materialized-view refreshes. The cap applies to each Postgres process, and parallel workers each get it, so keep cap × (workers + 1) well under free disk. It needs superuser or GRANT SET ON PARAMETER temp_file_limit.
  • Added postgresql.primary.extraArgs (server settings such as ["-c", "max_connections=200"]; changing them restarts the PostgreSQL pod), postgresql.primary.persistence.annotations and postgresql.primary.persistence.volumeAttributesClassName to the bundled PostgreSQL. Pod-spec only, no config.json change.
  • Added bifrost.relays (top-level relays): raw HTTP relays to non-LLM APIs such as Tavily, Exa or AWS Textract, served at /relay/{name}/.... Bifrost adds the relay’s credential, prices each counted call and charges it to the caller’s budgets. Entries render as written. With sourceOfTruth: config.json, an explicit list (including []) removes stored relays it doesn’t name, and their grants with them; leave the key out to keep stored relays. A repeated name fails the render. Grants are checked for identified callers only: with bifrost.client.enforceAuthOnInference: false, a request that carries no credential is relayed without any grant check.
  • Added relay_grants: [{relay_name: ...}] to bifrost.governance.virtualKeys[], bifrost.governance.projects[] and bifrost.accessProfiles[] (relay_grants), the relays a caller can reach. Like the rest of the entry, the list in your values is the whole list whenever that key, project or profile syncs from config.json: when its entry changes, or on every start with sourceOfTruth: config.json. Leaving relay_grants out then means no grants, so list any grant made in the dashboard that should stay. A grant naming a relay that doesn’t exist is skipped with a warning on a virtual key; a project or access profile naming one is skipped as a whole. A virtual key whose team, customer or user is governed by an access profile (Enterprise) takes its relays from that profile, and its own relay_grants are not applied; grant the relay on the profile.
  • Added podDisruptionBudget (top-level value): enabled: true keeps at least minAvailable (default 1) Bifrost pods running through node drains and cluster upgrades. Setting maxUnavailable replaces minAvailable, so maxUnavailable: "100%" lets every replica be evicted. Rendered only in cluster mode, and needs at least 2 replicas. No config.json change.
  • Added bifrost.scim.config.crossAppAccess (scim_config.config.crossAppAccess, every identity provider; Enterprise): lets AI clients your identity provider admits sign in to Bifrost as the user, by exchanging the provider’s ID-JAG for a Bifrost token. Takes enabled, accessTokenTtl (60-86400 seconds, default one day) and provisionUsers. It needs bifrost.client.oauth2ServerConfig.issuerUrl, the governance plugin, and an identity provider that issues ID-JAGs (such as Okta). Previously both the values schema and config.schema.json rejected it, so it could only be set from the dashboard.
  • Added bifrost.scim.config.tokenExchangeHosts (scim_config.config.tokenExchangeHosts, Enterprise): the hosts, besides the identity provider’s own, that an MCP server’s token exchange may be sent to. Previously both the values schema and config.schema.json rejected it, so it could only be set from the dashboard.
  • Added bifrost.scim.config.accessProfileResolutionStrategy (scim_config.config.accessProfileResolutionStrategy, Enterprise): order grants only the profiles of the first attribute rule a user matches; the default, all, grants every match. Previously both the values schema and config.schema.json rejected it, so it could only be set from the dashboard.
  • Added bifrost.scim.config.syncAllUsers (scim_config.config.syncAllUsers, Okta, default true): which users Okta directory sync keeps. true keeps every user assigned to the Okta app except suspended and deactivated ones; false keeps active users only. Directory sync runs only when apiToken is set. Previously the values schema rejected it, so it could only be set from the dashboard.
  • Added bifrost.client.skipAccessProfileUserVirtualKeyCreation and .skipAccessProfileUserVirtualKeyDeletion (client.skip_access_profile_user_virtual_key_creation / skip_access_profile_user_virtual_key_deletion, Enterprise): true stops giving users a virtual key with their first access profile, and keeps their virtual keys when their last profile is removed.
  • Added disable_content_logging to bifrost.governance.projects[] (governance.projects[].disable_content_logging): turns request and response content logging on or off for a project’s traffic, overriding the virtual key. Omit it to inherit.
  • Added the sagemaker provider (Amazon SageMaker real-time endpoints) to bifrost.providers (providers.sagemaker). Set sagemaker_key_config.region; leave out access_key and secret_key to use the default AWS credentials (for example IRSA). Models are endpoint names.
  • Added use_speech_endpoints and use_voice_live_endpoints to azure_key_config (providers.azure.keys[].azure_key_config): true sends speech and transcription to Azure AI Speech, and realtime sessions to Azure Voice Live. An alias can override either.
  • Added bifrost.providers.<name>.passthrough.allowed_routes (providers.<name>.passthrough.allowed_routes): limits which HTTP passthrough routes reach a provider, by path (* and /** globs) and optional methods. An empty list forwards nothing.
  • Added bifrost.warp (top-level warp): configures the dashboard’s Warp assistant (model and provider, limits, embeddings, log search and topics). Warp also needs bifrost.featureFlags.warp.enabled: true and a vector store.
  • Added bifrost.disableOnboarding (disable_onboarding): true hides the dashboard’s setup checklist for every user. It is read from config.json on every start, so it holds on stateless deployments.