* refactor: Restructure MCP registry system with caching
- Split MCPServersRegistry into modular components:
- MCPServerInspector: handles server inspection and health checks
- MCPServersInitializer: manages server initialization logic
- MCPServersRegistry: simplified registry coordination
- Add distributed caching layer:
- ServerConfigsCacheRedis: Redis-backed configuration cache
- ServerConfigsCacheInMemory: in-memory fallback cache
- RegistryStatusCache: distributed leader election state
- Add promise utilities (withTimeout) replacing Promise.race patterns
- Add comprehensive cache integration tests for all cache implementations
- Remove unused MCPManager.getAllToolFunctions method
* fix: Update OAuth flow to include user-specific headers
* chore: Update Jest configuration to ignore additional test files
- Added patterns to ignore files ending with .helper.ts and .helper.d.ts in testPathIgnorePatterns for cleaner test runs.
* fix: oauth headers in callback
* chore: Update Jest testPathIgnorePatterns to exclude helper files
- Modified testPathIgnorePatterns in package.json to ignore files ending with .helper.ts and .helper.d.ts for cleaner test execution.
* ci: update test mocks
---------
Co-authored-by: Danny Avila <danny@librechat.ai>
* 📫 refactor: Enhance OpenID email Fallback
* Updated email retrieval logic to use preferred_username or upn if email is not available.
* Adjusted logging and user data assignment to reflect the new email handling approach.
* Ensured email domain validation checks the correct email source.
* 🔄 refactor: Update Email Domain Validation Logic
* Modified `isEmailDomainAllowed` function to return true for falsy emails and missing domain restrictions.
* Added new test cases to cover scenarios with and without domain restrictions.
* Ensured proper validation when domain restrictions are present.
* fix: Sanitize LLM titles by stripping <think> tags and fix modal overflow
* chore: linting
* chore: Simplify title sanitization by removing unnecessary variable assignment and import order
---------
Co-authored-by: Danny Avila <danny@librechat.ai>
* 📦 feat: `@librechat/agents` v2.4.87 for LangFuse Support
* 📦 chore: update @librechat/agents to v2.4.88 in package.json and package-lock.json
* 📦 chore: update @librechat/agents to v2.4.89
* feat: Add runName configuration to AgentClient and Memory agent for improved tracing
* updated gpt5-pro
it is here and on openrouter
https://platform.openai.com/docs/models/gpt-5-pro
* feat: Add gpt-5-pro pricing
- Implemented handling for the new gpt-5-pro model in the getValueKey function.
- Updated tests to ensure correct behavior for gpt-5-pro across various scenarios.
- Adjusted token limits and multipliers for gpt-5-pro in the tokens utility files.
- Enhanced model matching functionality to include gpt-5-pro variations.
* refactor: optimize model pricing and validation logic
- Added new model pricing entries for llama2, llama3, and qwen variants in tx.js.
- Updated tokenValues to include additional models and their pricing structures.
- Implemented validation tests in tx.spec.js to ensure all models resolve correctly to pricing.
- Refactored getValueKey function to improve model matching and resolution efficiency.
- Removed outdated model entries from tokens.ts to streamline pricing management.
* fix: add missing pricing
* chore: update model pricing for qwen and gemma variants
* chore: update model pricing and add validation for context windows
- Removed outdated model entries from tx.js and updated tokenValues with new models.
- Added a test in tx.spec.js to ensure all models with pricing have corresponding context windows defined in tokens.ts.
- Introduced 'command-text' model pricing in tokens.ts to maintain consistency across model definitions.
* chore: update model names and pricing for AI21 and Amazon models
- Refactored model names in tx.js for AI21 and Amazon models to remove versioning and improve consistency.
- Updated pricing values in tokens.ts to reflect the new model names.
- Added comprehensive tests in tx.spec.js to validate pricing for both short and full model names across AI21 and Amazon models.
* feat: add pricing and validation for Claude Haiku 4.5 model
* chore: increase default max context tokens to 18000 for agents
* feat: add Qwen3 model pricing and validation tests
* chore: reorganize and update Qwen model pricing in tx.js and tokens.ts
---------
Co-authored-by: khfung <68192841+khfung@users.noreply.github.com>
* fix: Remove ephemeral cache control from document encoding function
* refactor: Improve document encoding types and add file context for anthropic messages api
- Added AnthropicDocumentBlock interface to define the structure for documents from the Anthropic provider.
- Updated encodeAndFormatDocuments function to utilize the new type and include optional context for filenames.
- Refactored DocumentResult to use a union type for various document formats, improving type safety and clarity.
- problem: `addImageUrls` had a side effect that was being leveraged before to populate both the `ocr` message field, now `fileContext`, and `client.options.attachments`, which would record the user's uploaded message attachments to the user message when saved to the database and returned at the end of the request lifecycle
- solution: created dedicated handling for file context, and made sure to populate `allFiles` with non-provider attachments
* 📎 feat: Direct Provider Attachment Support for Multimodal Content
* 📑 feat: Anthropic Direct Provider Upload (#9072)
* feat: implement Anthropic native PDF support with document preservation
- Add comprehensive debug logging throughout PDF processing pipeline
- Refactor attachment processing to separate image and document handling
- Create distinct addImageURLs(), addDocuments(), and processAttachments() methods
- Fix critical bugs in stream handling and parameter passing
- Add streamToBuffer utility for proper stream-to-buffer conversion
- Remove api/agents submodule from repository
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* chore: remove out of scope formatting changes
* fix: stop duplication of file in chat on end of response stream
* chore: bring back file search and ocr options
* chore: localize upload to provider string in file menu
* refactor: change createMenuItems args to fit new pattern introduced by anthropic-native-pdf-support
* feat: add cache point for pdfs processed by anthropic endpoint since they are unlikely to change and should benefit from caching
* feat: combine Upload Image into Upload to Provider since they both perform direct upload and change provider upload icon to reflect multimodal upload
* feat: add citations support according to docs
* refactor: remove redundant 'document' check since documents are handled properly by formatMessage in the agents repo now
* refactor: change upload logic so anthropic endpoint isn't exempted from normal upload path using Agents for consistency with the rest of the upload logic
* fix: include width and height in return from uploadLocalFile so images are correctly identified when going through an AgentUpload in addImageURLs
* chore: remove client specific handling since the direct provider stuff is handled by the agent client
* feat: handle documents in AgentClient so no need for change to agents repo
* chore: removed unused changes
* chore: remove auto generated comments from OG commit
* feat: add logic for agents to use direct to provider uploads if supported (currently just anthropic)
* fix: reintroduce role check to fix render error because of undefined value for Content Part
* fix: actually fix render bug by using proper isCreatedByUser check and making sure our mutation of formattedMessage.content is consistent
---------
Co-authored-by: Andres Restrepo <andres@thelinuxkid.com>
Co-authored-by: Claude <noreply@anthropic.com>
📁 feat: Send Attachments Directly to Provider (OpenAI) (#9098)
* refactor: change references from direct upload to direct attach to better reflect functionality
since we are just using base64 encoding strategy now rather than Files/File API for sending our attachments directly to the provider, the upload nomenclature no longer makes sense. direct_attach better describes the different methods of sending attachments to providers anyways even if we later introduce direct upload support
* feat: add upload to provider option for openai (and agent) ui
* chore: move anthropic pdf validator over to packages/api
* feat: simple pdf validation according to openai docs
* feat: add provider agnostic validatePdf logic to start handling multiple endpoints
* feat: add handling for openai specific documentPart formatting
* refactor: move require statement to proper place at top of file
* chore: add in openAI endpoint for the rest of the document handling logic
* feat: add direct attach support for azureOpenAI endpoint and agents
* feat: add pdf validation for azureOpenAI endpoint
* refactor: unify all the endpoint checks with isDocumentSupportedEndpoint
* refactor: consolidate Upload to Provider vs Upload image logic for clarity
* refactor: remove anthropic from anthropic_multimodal fileType since we support multiple providers now
🗂️ feat: Send Attachments Directly to Provider (Google) (#9100)
* feat: add validation for google PDFs and add google endpoint as a document supporting endpoint
* feat: add proper pdf formatting for google endpoints (requires PR #14 in agents)
* feat: add multimodal support for google endpoint attachments
* feat: add audio file svg
* fix: refactor attachments logic so multi-attachment messages work properly
* feat: add video file svg
* fix: allows for followup questions of uploaded multimodal attachments
* fix: remove incorrect final message filtering that was breaking Attachment component rendering
fix: manualy rename 'documents' to 'Documents' in git since it wasn't picked up due to case insensitivity in dir name
fix: add logic so filepicker for a google agent has proper filetype filtering
🛫 refactor: Move Encoding Logic to packages/api (#9182)
* refactor: move audio encode over to TS
* refactor: audio encoding now functional in LC again
* refactor: move video encode over to TS
* refactor: move document encode over to TS
* refactor: video encoding now functional in LC again
* refactor: document encoding now functional in LC again
* fix: extend file type options in AttachFileMenu to include 'google_multimodal' and update dependency array to include agent?.provider
* feat: only accept pdfs if responses api is enabled for openai convos
chore: address ESLint comments
chore: add missing audio mimetype
* fix: type safety for message content parts and improve null handling
* chore: reorder AttachFileMenuProps for consistency and clarity
* chore: import order in AttachFileMenu
* fix: improve null handling for text parts in parseTextParts function
* fix: remove no longer used unsupported capability error message for file uploads
* fix: OpenAI Direct File Attachment Format
* fix: update encodeAndFormatDocuments to support OpenAI responses API and enhance document result types
* refactor: broaden providers supported for documents
* feat: enhance DragDrop context and modal to support document uploads based on provider capabilities
* fix: reorder import statements for consistency in video encoding module
---------
Co-authored-by: Dustin Healy <54083382+dustinhealy@users.noreply.github.com>
* 🔧 chore: Update @librechat/agents to v2.4.84 in package.json and package-lock.json
* feat: Serper as new scraperProvider for Web Search and add firecrawlVersion support
* fix: TWebSearchKeys and ensure unique API keys extraction
* chore: Add build:packages script to streamline package builds
* fix: update @librechat/agents to v2.4.83 to handle reasoning edge case encountered with GLM models
* feat: GLM Context Window & Pricing Support
* feat: Add support for glm4 model in token values and tests
* chore: linting for `loadCustomConfig`
* refactor: decouple CDN init and variable/health checks from AppService
* refactor: move AppService to packages/data-schemas
* chore: update AppConfig import path to use data-schemas
* chore: update JsonSchemaType import path to use data-schemas
* refactor: update UserController to import webSearchKeys and redefine FunctionTool typedef
* chore: remove AppService.js
* refactor: update AppConfig interface to use Partial<TCustomConfig> and make paths and fileStrategies optional
* refactor: update checkConfig function to accept Partial<TCustomConfig>
* chore: fix types
* refactor: move handleRateLimits to startup checks as is an effect
* test: remove outdated rate limit tests from AppService.spec and add new handleRateLimits tests in checks.spec
* ⚙️ chore: Resolve Build Warning and `keyvMongo` types
* 🔄 chore: Update mongodb version to ^6.14.2 in package.json and package-lock.json
* chore: remove @langchain/openai dep
* 🔄 refactor: Change log level from warn to debug for missing endpoint config
* 🔄 refactor: Improve temp chat expiration date calculation in tests and implementation
* initialize servers sequentially
* adjust for exported properties that are not nullable anymore
* use underscore separator
* mock with set
* customize init timeout via env var
* refactor for readability, use loaded conns for tool functions
* address PR comments
* clean up fire-and-forget
* fix tests
* Refactor: Moved Redis cache infra logic into `packages/api`
- Moved cacheFactory and redisClients from `api/cache` into `packages/api/src/cache` so that features in `packages/api` can use cache without importing backward from the backend.
- Converted all moved files into TS with proper typing.
- Created integration tests to run against actual Redis servers for redisClients and cacheFactory.
- Added a GitHub workflow to run integration tests for the cache feature.
- Bug fix: keyvRedisClient now implements the PING feature properly.
* chore: consolidate imports in getLogStores.js
* chore: reorder imports
* chore: re-add fs-extra as dev dep.
* chore: reorder imports in cacheConfig.ts, cacheFactory.ts, and keyvMongo.ts
---------
Co-authored-by: Danny Avila <danny@librechat.ai>
* ✨ v0.8.0
* 🔧 chore: Update config version to 1.3.0
* 🔧 chore: Bump @librechat/api version to 1.4.1
* 🔧 chore: Update @librechat/client version to 0.3.1
* 🔧 chore: Bump librechat-data-provider version to 0.8.020
* 🔧 chore: Bump @librechat/data-schemas version to 0.0.23
* initialize servers sequentially
* adjust for exported properties that are not nullable anymore
* use underscore separator
* mock with set
* customize init timeout via env var
* feat: Add new Claude models to sharedAnthropicModels list
* chore: use correct claude aliases for default list
* chore: update deepseek model rates for accuracy
* chore: update @librechat/agents dependency to version 2.4.82
* chore: move domain related functions to `packages/api`
* fix: isEmailDomainAllowed for case-insensitive domain matching
- Added tests to validate case-insensitive matching for email domains in various scenarios.
- Updated isEmailDomainAllowed function to convert email domains to lowercase for consistent comparison.
- Improved handling of null/undefined entries in allowedDomains.
* ci: Mock isEmailDomainAllowed in samlStrategy tests
- Added a mock implementation for isEmailDomainAllowed to return true in samlStrategy tests, ensuring consistent behavior during test execution.
* ci: Update import of isEmailDomainAllowed in ldapStrategy tests
- Changed the import of isEmailDomainAllowed from the domains service to the api package for consistency and to reflect recent refactoring.
* refactor: move `loadOCRConfig` from `packages/data-provider` to `packages/api` and return `undefined` if not explicitly configured
* fix: loadOCRConfig import from @librechat/api
* refactor: update defaultTextMimeTypes to support virtually all file types for text parsing
* fix: improve OCR capability check and error message for unsupported file types
* ci: remove unnecessary ocr expectation from AppService test
* 🌐 feat: OpenRouter Web Search
- Added tests for handling web_search parameter with OpenRouter in various scenarios.
- Implemented logic to manage web_search in modelOptions and addParams/dropParams.
- Ensured correct configuration of llmConfig and modelKwargs for OpenRouter, including handling of plugins.
- Improved overall integration of OpenRouter with OpenAI API, ensuring expected behavior across different configurations.
* chore: bump @librechat/agents to v2.4.81
- Added tests to validate behavior of web_search parameter in getOpenAIConfig function.
- Implemented logic to handle web_search in addParams and dropParams, ensuring correct precedence and behavior.
- Ensured web_search does not appear in modelKwargs or llmConfig when not applicable.
- Improved overall configuration management for OpenAI API integration.
- Fixed a bug in reinitMCPServer where a user connection was created for an app-level server whenever this server is reinitialized
- Made MCPManager.getUserConnection to return an error if the connection is app-level
- Add MCPManager.getConnection to return either an app connection or a user connection based on the serverName
- Made MCPManager.appConnections public to avoid unnecessary wrapper methods.
* add use of immutable claims to identify user object
* fix semicolons
* update email attribute on change
* replace ternary expressions
* fix semicolon
* chore: add typing
* chore: reorder fields in `findOpenIDUser`
* refactor: optimize user lookup logic in `findOpenIDUser` function to minimize database roundtrips
* refactor: integrate findOpenIDUser for improved user retrieval in refreshController
* refactor: improve error logging for invalid refresh tokens in refreshController
* ci: mock findUser correctly in openidStrategy tests
* test: add unit tests for findOpenIDUser function to enhance user retrieval logic
---------
Co-authored-by: Joachim Keltsch <joachim.keltsch@daimlertruck.com>
* chore: Improve logging format for initial flow state creation
* refactor: MCP Tool Management with Improved Auth Handling and State Management
- Updated `CustomUserVarsSection` to include optional localization for placeholder text.
- Refactored `MCPToolSelectDialog` to streamline tool addition and management, including improved handling of authentication data.
- Introduced a new `addToolsToForm` function to encapsulate logic for adding tools to the form state.
- Enhanced `useRemoveMCPTool` hook to simplify tool removal logic and ensure proper state updates.
- Added loading state management for custom variable saving to improve user experience.
* refactor: Enhance MCP Tool Removal Logic and Integrate Toast Notifications
- Updated `MCPToolSelectDialog` to utilize the new `removeTool` function from the `useRemoveMCPTool` hook for improved tool removal handling.
- Refactored `useRemoveMCPTool` to accept options for toast notifications, allowing for more flexible user feedback during tool removal.
- Removed the previous inline tool removal logic to streamline the component's code and improve maintainability.
* refactor: Enhance user plugins mutation to invalidate MCP auth values on uninstall
* refactor: Replace refetchQueries with invalidateQueries for improved cache management
* chore: remove unused i18n key
* refactor: Implement gradual backoff polling for oauth connection status with timeout handling
* refactor: Enhance OAuth polling with gradual backoff and timeout handling; update reconnection tracking
* refactor: reconnection timeout behavior in OAuthReconnectionManager and OAuthReconnectionTracker
- Implement tests to verify reconnection timeout handling, including tracking of reconnection states and cleanup of timed-out entries.
- Enhance existing methods in OAuthReconnectionManager and OAuthReconnectionTracker to support timeout checks and cleanup logic.
- Ensure proper handling of multiple servers with different timeout periods and edge cases for active states.
* chore: remove comment
* refactor: Enforce strict 3-minute OAuth timeout with updated polling intervals and improved timeout handling
* refactor: Remove unused polling logic and prevent duplicate polling for servers in MCP server manager
* refactor: Update localization key for no memories message in MemoryViewer
* refactor: Improve MCP tool initialization by handling server failures
- Introduced a mechanism to track failed MCP servers, preventing retries for unavailable servers.
- Added logging for failed tool creation attempts to enhance debugging and monitoring.
* refactor: Update reconnection timeout to enforce a strict 3-minute limit
* ci: Update reconnection timeout tests to reflect a strict 3-minute limit
* ci: Update reconnection timeout tests to enforce a strict 3-minute limit
* chore: Remove unused MCP connection timeout message
* fix: axios response logging for text parsing, remove console logging, remove jsdoc
* refactor: error logging in logAxiosError function to handle various error types with type guards
* refactor: enhance text parsing with improved error handling and async file reading
* refactor: replace synchronous file reading with asynchronous methods for improved performance and memory management
* ci: update tests
* chore: Update documentation for formatToolContent function, remove JSDoc types and duplicate comments
* chore: fix type errors due to attachment.filename in Attachment component
* fix: respect server's token endpoint auth methods for MCP OAuth refresh
Previously, LibreChat always used Basic Auth when refreshing OAuth tokens if a
client_secret was present. This caused issues with servers (like FastMCP) that
only support client_secret_post. Now properly checks and respects the server's
advertised token_endpoint_auth_methods_supported.
Fixes token refresh failures with error: "refresh_token.client_id: Field required"
* chore: remove MCP OAuth URL Logging
- Updated MCPServersRegistry to correctly process serverInstructions when provided as a string "true", allowing it to fetch instructions from the server.
- Added a new utility function `isEnabled` to determine if serverInstructions should trigger a fetch.
- Introduced comprehensive tests to validate the behavior for different serverInstructions configurations, ensuring both string "true" and boolean true fetch from the server while custom strings remain unchanged.
🎯 This enhancement improves the flexibility and correctness of server instruction handling in the MCPServersRegistry.
* WIP: conversion of `ocr` to `context`
* refactor: make `primeResources` backwards-compatible for `ocr` tool_resources
* refactor: Convert legacy `ocr` tool resource to `context` in agent updates
- Implemented conversion logic to replace `ocr` with `context` in both incoming updates and existing agent data.
- Merged file IDs and files from `ocr` into `context` while ensuring deduplication.
- Updated tools array to reflect the change from `ocr` to `context`.
* refactor: Enhance context file handling in agent processing
- Updated the logic for managing context files by consolidating file IDs from both `ocr` and `context` resources.
- Improved backwards compatibility by ensuring that context files are correctly populated and handled.
- Simplified the iteration over context files for better readability and maintainability.
* refactor: Enhance tool_resources handling in primeResources
- Added tests to verify the deletion behavior of tool_resources fields, ensuring original objects remain unchanged.
- Implemented logic to delete `ocr` and `context` fields after fetching and re-categorizing files.
- Preserved context field when the context capability is disabled, ensuring correct behavior in various scenarios.
* refactor: Replace `ocrEnabled` with `contextEnabled` in AgentConfig
* refactor: Adjust legacy tool handling order for improved clarity
* refactor: Implement OCR to context conversion functions and remove original conversion logic in update agent handling
* refactor: Move contextEnabled declaration to maintain consistent order in capabilities
* refactor: Update localization keys for file context to improve clarity and accuracy
* chore: Update localization key for file context information to improve clarity
