Changelog
This is a major release that introduces deny-by-default semantics across all allow-list fields (models, keys, tools, providers), a dedicated Provider Keys API, blacklist support in load balancing, redesigned adaptive routing UI, and scoped pricing overrides. This release contains multiple breaking changes — please review the breaking changes section and migration checklist carefully before upgrading.⚠️ Breaking Changes
v1.5.0 OSS base flips the meaning of empty arrays across all allow-list fields. Existing deployments with a database are protected by automatic migrations on startup, but any new configuration created after upgrading must follow the new semantics. Back up your config store database before upgrading — this migration is not revertible.
| What you write | v1.4.x meaning | v1.5.0 meaning |
|---|---|---|
[] (empty array) | Allow all | Allow none (deny by default) |
["*"] (wildcard) | N/A | Allow all |
["a", "b"] | Only a and b | Only a and b (unchanged) |
1. Provider Key models Field
Empty models array now means “allow none” instead of “allow all”. Use ["*"] to allow a key to serve all models.2. Virtual Key allowed_models Field
Missing or empty allowed_models on a VK provider config now blocks all models from that provider. Use ["*"] to allow all.3. Virtual Key Provider Configs — Deny-by-Default
Virtual Keys with empty or missingprovider_configs now block all providers. Every VK must explicitly list its permitted providers.4. allowed_keys Renamed to key_ids
Field renamed in VK provider configs. Same deny-by-default semantics — omitted or empty key_ids now blocks all keys. Use ["*"] to allow all. Note: Unlike allowed_models, there is no automatic migration for key_ids.5. Virtual Key MCP tools_to_execute Field
Empty tools_to_execute now blocks all tools. The mcp_configs list itself acts as a strict allow-list — no mcp_configs means all MCP tools are blocked for that VK.6. weight Field is Now Optional
weight on VK provider configs is now nullable (*float64). null or omitted means the provider is excluded from weighted routing but still reachable via direct routing or fallbacks.7. Compat Plugin Configuration Changes
enable_litellm_fallbacksoption removed- Replaced with:
compat.convert_text_to_chat,compat.convert_chat_to_responses,compat.should_drop_params - Response field
extra_fields.litellm_compatremoved - New response fields:
extra_fields.dropped_compat_plugin_params,extra_fields.converted_request_type
8. Image Edits No Longer Supported on Replicate’s Image Generation Endpoint
/v1/images/generations on Replicate now only handles pure text-to-image generation. Image editing parameters must use /v1/images/edits. Note: /v1/images/edits on Replicate will also be removed in a follow-up release.9. Provider Keys API Separated from Provider API
keysfield removed from provider create/update requests and responses- New dedicated endpoints:
GET/POST /api/providers/{provider}/keys,GET/PUT/DELETE /api/providers/{provider}/keys/{key_id} - Create providers first, then add keys separately
New Validation: WhiteList Rules
- Wildcard
["*"]cannot be mixed with other values (HTTP 400) - No duplicate values allowed in allow-list fields
- Applies to:
allowed_models,key_ids,models,tools_to_execute,tools_to_auto_execute,allowed_extra_headers
Quick Migration Checklist
- Update provider key
modelsin config.json — change[]to["*"] - Add
allowed_models: ["*"]to every VK provider config - Ensure every VK has at least one provider config entry
- Rename
allowed_keystokey_idsand set["*"]where needed - Update
tools_to_executefor MCP configs — change[]to["*"] - Handle nullable
weightin API consumers - Fix any invalid WhiteList values (no mixing wildcards, no duplicates)
- Migrate key management to dedicated
/api/providers/{provider}/keysendpoints
✨ Features
- Dedicated Provider Keys API — Keys are now managed via
/api/providers/{provider}/keysendpoints instead of being embedded in provider create/update payloads - Deny-by-Default Access Control — Standardized empty array conventions across all allow-list fields;
[]means deny all,["*"]means allow all - VK Provider Config Key Wildcards —
key_idsnow supports["*"]wildcard to allow all keys; handler resolves wildcard to AllowAllKeys flag without DB key lookups - VK MCP Allow-List — Virtual key MCP configs now act as an execution-time allow-list — tools not permitted by the VK are blocked at inference and MCP tool execution
- MCP Virtual Key Assignment — MCP configuration now supports assigning virtual keys with per-tool access control, with an option to allow MCP clients to run on all virtual keys
- Disable Auto MCP Tool Injection — Add option to disable automatic MCP tool injection per request
- MCP Request-Level Extra Headers — Support for request-level extra headers in MCP tool execution
- MCP Gateway Filtering — Support for
x-bf-mcp-include-clientsandx-bf-mcp-include-toolsrequest headers to filter MCP tools/list response - Scoped Pricing Overrides — Support for pricing overrides at a scoped level
- StabilityAI on Bedrock — Added StabilityAI provider support to Bedrock
- Plugin Trace Logging — Plugins can now inject logs at trace level using
ctx.Log(schemas.LogLevelInfo, "Test log") - Blacklist Support in Load Balancing — Added model blacklist support to the load balancing plugin
- Adaptive Routing UI Redesign — Redesigned adaptive routing UI with improved layout and Sankey chart visualization
- Governance Refactor — Governance module changes for improved structure
- Compat Plugin New Modes — Chat-to-responses fallback and OpenAI-compatible parameter dropping modes added to compat plugin
🐞 Fixed
- MCP Agent Usage Accumulation — Fixed accumulated usage not being sent back in MCP agent mode
- OpenAI Transcription Formats — Handle text, vtt, srt response formats in OpenAI transcription response
- HuggingFace Load Balancing — Removed HuggingFace deployment handling from load balancing plugin
- Parallelized Model Listing — Parallelized model listing for providers to speed up startup time
📀 Base OSS version
transports/v1.5.0-prerelease1