AnythingLLM changelog: what's new each month
Every stable AnythingLLM release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.
November 2026 · so far
1 release: 1.20.0
November brought major improvements to presentation creation with dynamic themes and live previews, plus significant enhancements to MCP server management and developer capabilities.
Highlights
- Slides now preview live in chat as they build, so you can see the deck come together in real time.
- Automatic theme and layout selection ensures each slide fits its content without repetition.
- MCP servers can now forward workspace and user identity, and developers have full API control over server management.
New
- Live slide previews appear in chat and can be clicked to view full size.
- MCP servers support identity forwarding to receive workspace and user context.
- Per-server timeout configuration allows control over tool call duration.
- Developer API endpoints for complete MCP server management (list, add, update, delete, start, stop, reload).
- Individual tool on/off control through the API for MCP servers.
Improved
- Presentation themes and layouts now automatically adapt to each slide's content.
- Chart values, figures and quotes are grounded only when they appear in source material.
- Text sizing automatically fits content to slides.
- MCP servers with many tools now load all pages of tools.
- Starting a server picks up any newly saved servers.
- Markdown tables now scroll properly on narrow screens.
Fixed
- Markdown tables on narrow screens are now wrapped in a scroll container.
- Disconnected MCP servers can now be deleted.
October 2026
3 releases: 1.17.0 → 1.19.0
October brought stability improvements and fixes across data handling, agent interactions, and provider integrations. The month focused on resolving edge cases in file parsing, web browsing, and API compatibility.
- Temporary parameter dropped for Anthropic models.
- Open-computer tool has been moved to its own repository.
- Weaviate upgraded to weaviate-client v3.
Highlights
- Agent web browsing now supports SearchApi engines.
- Spreadsheet dates are now read correctly instead of as serial numbers.
- Images in attachments now open when clicked, even with duplicate file names.
- Responses now show a notice when hitting max output token limits.
- Added support for .mov files in transcription.
New
- Support for .mov files in transcription.
- Notice displayed when response hits max output token limit.
- Dismissal comment feature to prevent low-effort research claims.
- Bedrock India inference profiles now route via bedrock-runtime.
- Clarifying question timeout now holds once the user starts answering.
Improved
- HTML-only mbox messages are now read instead of skipped.
- Text uploads are read in their original encoding.
- Weaviate and Astra now score chunks by cosine similarity like other providers.
- Kubernetes probes now point at the server's /api/ping endpoint.
- Agent tool required parameters are sent on every provider path.
- SDK retry count is now honored when ANYTHINGLLM_MAX_RETRIES is set.
- Spreadsheet formulas are neutralized in CSV chat exports.
Fixed
- Server no longer crashes on malformed agent WebSocket frames.
- Generic OpenAI now supports max tokens of 0 to omit max_tokens from request.
- Fixed crash when removing a parsed file while an image is attached.
- History attachments are no longer sent to providers that build prompts from raw chat history.
- Embed allowlist entries are now compared as browser origins.
- Unauthenticated embed DoS prevented by capping request body size.
September 2026
1 release: 1.16.2
September brought significant expansions to AnythingLLM's capabilities, including new AI providers, agent skills, and language support. The month focused on making local model management easier and improving web search functionality.
- Docker Model Runner replaced by LLMman OSS
Highlights
- LLMman OSS is now available as a local LLM provider, replacing Docker Model Runner
- Google Vertex AI added as an LLM provider option
- New generate-image agent skill allows creating images directly from chat
- You.com is now the default web search provider with Zero Data Retention
- Hindi, Bulgarian, and Swedish language support added
New
- LLMman OSS local LLM provider
- Google Vertex AI LLM provider
- Generate-image agent skill for image creation
- You.com as default web search provider
- Agent flow variables with optional arguments support
- Hindi, Bulgarian, and Swedish language support
- GenieX on Windows ARM now shows local model directory and allows pulling from Hugging Face UI
- Version tag click shortcut to logs and storage directory
Improved
- Web search now uses You.com for richer results with Zero Data Retention
- GenieX on Windows ARM can discover and manage models directly from the UI
- Easier access to logs and storage directory through version tag in window header
August 2026
2 releases: 1.16.0 → 1.16.1
August brought improvements to model support across multiple providers and refinements to the user interface. Key additions include native Windows SDK support for Foundry Local and expanded compatibility with various AI services.
Highlights
- Foundry Local is now fully self-contained on Windows without requiring a separate installation.
- AWS Bedrock now supports cross-region inference profiles with automatic routing.
- LocalAI added as an image generation provider with automatic context window detection.
- Chain of thought and agent activity combined into a single collapsible interface component.
- OpenAI-compatible Whisper API support added for speech recognition.
New
- Native Foundry Local SDK for Windows x64 and ARM64 with GPU/NPU optimization.
- GenieX runtime support for Windows ARM64 devices.
- LocalAI as an image generation provider.
- OpenAI-compatible Whisper API support.
- Silent and headless install/uninstall flags for Windows installer.
- LocalAI context window auto-detection from server configuration.
Improved
- Username character limit increased to 64.
- Default fallback models updated for OpenAI and Anthropic.
- Chain of thought and agent activity display now uses a single collapsible component with visual rail connections.
- Assistant Panel now mirrors the new chain of thought rollup interface.
- Kokoro-TTS voice list rendering for pre-0.3.x kokoro-fastapi servers.
- AWS Bedrock profile dropdown now shows geo-prefixed profile IDs for easier region selection.
Fixed
- Temperature setting disabled for Claude Sonnet 5 model.
- Anthropic non-streaming chat completions now stream correctly in REST API workspace chat.
- Kokoro-TTS voice list rendering for older server versions.
- Catalan translation typos and mistakes corrected.
June 2026
4 releases: 1.14.0 → 1.15.0
June brought significant enhancements to AnythingLLM's agent capabilities, audio processing, and multi-platform support. The month featured new AI providers, improved tool calling, and major performance optimizations across GPU support and binary sizes.
- Default chat thread behavior changed: new threads created will close the default thread if it contains chats
- Tool calling now opt-out by default instead of opt-in; users must explicitly disable if not wanted
Highlights
- Added support for Intel, AMD, and NVIDIA GPUs with 92% smaller binary size and 15% faster processing
- Introduced Cerebres provider and multiple new speech-to-text and text-to-speech options
- Enhanced Meeting Assistant with better speaker identification, dual-channel stereo support, and improved context handling for longer meetings
- New web search providers (Brave Search, fastCRW) and improved web scraping with markdown conversion
- Tool calling now enabled by default across all model providers for better agent performance
New
- Cerebres provider support
- Speech-to-text support for Deepgram, GenericOAI, Lemonade, and OpenAI
- Text-to-speech support for KokoroTTS
- Developer API for audio transcription
- Basic Speaker Identification for meeting summaries
- Dual channel stereo recording support for meetings
- Brave Search and fastCRW web search providers
- Intelligent Tool Selection (ToolReranker) for better agent decisions
Improved
- Web scraping now converts to markdown for better parsing and reduced context usage
- Summary tool overhauled with better transparency and confirmation prompts for longer summaries
- GenericOAI provider improvements
- Meeting Assistant context window overflow handling for better small model performance
- Speaker identification accuracy improved 80% in Full Diarization mode
- Linux AppImage 91% smaller with Ollama engine download caching
- Meeting titles now auto-update in UI after summarization
- Chat history cap unlocked for extended conversations
Fixed
- Edit prompt not appearing on user prompts
- Attached file context and citations in empty workspaces and agent chat
- API update-embeddings failure on Windows paths with Prisma
- Missing headers in Ollama embedder engine isAlive method
- Meeting Assistant title updates now reflected immediately in the interface
May 2026
1 release: 1.13.0
May brought hybrid AI routing and intelligent agent scheduling to AnythingLLM. Users can now mix local and cloud models in conversations, and automate agent tasks on custom schedules without code.
Highlights
- Mix local models (Ollama, LM Studio) and cloud providers (OpenAI, Anthropic, Google) in a single conversation with automatic switching.
- Schedule agents to run automatically at specific times or intervals using a visual interface with no cron syntax required.
- Route queries intelligently to cheaper or local models for simple tasks, saving costs while reserving expensive APIs for complex requests.
- Create custom routing rules based on keywords, token counts, time of day, image attachments, or LLM-classified intent.
- Sticky routing keeps conversations on the same model to avoid switching between providers on each message.
New
- Hybrid AI mode lets you combine multiple model providers in one conversation.
- Visual Cron Builder for scheduling agent tasks with dropdown menus instead of command-line syntax.
- Intelligent caching and sticky routing to maintain model consistency within conversation threads.
- LLM-classified rules that understand message intent for smarter routing decisions.
- Complete run history logging with full reasoning, tool calls, generated files, and agent responses.
- Ability to continue interrupted agent runs from previous executions.
Improved
- Scheduled jobs now have access to all agent tools including web search, document analysis, custom skills, and MCP integrations.
- Cost optimization through automatic model selection based on query complexity.
- Rule-based routing now supports time-of-day and attachment-type conditions for more granular control.
April 2026
2 releases: 1.12.0 → 1.12.1
April brought major expansions to AnythingLLM's agent capabilities and multi-modal support. The month focused on deepening integration features and improving document handling across languages.
- Agent skill auto-approval now requires ENV setting to enable (previously default behavior)
Highlights
- Full agent skill system now available within AnythingLLM, with automatic mode and @agent support for seamless workflow integration.
- Multi-modal chat expanded with image understanding, voice messages, and file attachments alongside text.
- Enhanced document generation with native application opening and improved PDF support for Asian languages.
- Google Calendar integration added for agent skills.
- Citations system improved for both direct chat and app integrations.
New
- Image understanding capability in chat
- Voice message support
- File attachment support
- Automatic agent mode with @agent mention syntax
- Workspace and thread selection controls
- Per-message model selection
- Google Calendar integration
- Image lightbox viewer in main UI
Improved
- Text chat now supports streaming responses and AI thinking display
- Citations system improved for app integrations
- PDF generation now supports Korean, Chinese, and Japanese characters
- Document Generation Agent opens generated files in native applications
- Web search in agent skills now defaults to DuckDuckGo
- Ollama updated to 0.20.7 with support for Qwen 3.5, Gemma 4, and other new models
March 2026
2 releases: 1.11.1 → 1.11.2
March brought significant improvements to AnythingLLM's agent capabilities, with enhanced tool calling, better citation tracking, and more control over agent skills. The month focused on making agents more powerful and easier to manage.
Highlights
- Agents now support complex multi-step tool calls with improved response quality
- New metrics and reporting for agent calls with better citation tracking
- Toggle agent skills on and off directly from the prompt interface
- Select workspace provider and model without leaving the page
New
- Complex multi-step tool calls for LLM agents
- New prompt input interface
- Metrics reporting for agent calls including finish reason
- Citation reporting during agent calls from documents and web searches
- Ability to toggle agent skills on or off from the prompt
Improved
- Agent models now continue working until final response is generated
- Citations UI and reporting redesigned
- Provider and model selection streamlined in workspace settings
- Tool-calling model responses improved significantly
Fixed
- Yarn setup:envs no longer fails when .env files already exist
February 2026
1 release: 1.11.0
February brought interface refinements and usability improvements to AnythingLLM, including a redesigned login experience and new keyboard navigation options. The release focused on standardizing user management and improving local model handling.
Highlights
- New login page UI redesign for improved user experience.
- Keyboard shortcuts added to quickly jump to the top or bottom of chat history.
- Username constraints standardized across the platform.
- Ollama context preferences now persist within language chain tools.
New
- Keyboard shortcuts to scroll to top and bottom of chat history.
Improved
- Username constraints refined and standardized.
- Ollama context window setting refactored.
- Ollama context preferences now persist in LC tools.
- YouTube file collector API patched.
Fixed
- YouTube file collector API issue resolved.
January 2026
1 release: 1.10.0
January brought major new capabilities for meeting recording and file processing, all running locally on your device without external dependencies or usage limits.
Highlights
- Record and summarize meetings from any platform (Zoom, Slack, Discord, Teams) without joining them.
- Process and chat with arbitrary files using on-device AI powered by NVIDIA Parakeet.
- Use custom summary templates and speaker identification for recorded content.
- Access all your existing AnythingLLM tools and integrations within meeting recordings.
- Removed rate limits and usage caps for local processing.
New
- Meeting recording that works across any meeting software.
- Passive meeting monitoring with desktop notifications to start recording.
- Speaker identification for recorded meetings.
- Custom summary templates for meeting transcripts.
- Chat interface for interacting with meeting transcripts.
- On-device file summarization for arbitrary files.
Improved
- Onboarding flow now skips workspace creation and goes directly to home.
- Workspace file picker performance improved.
December 2025
1 release: 1.9.1
December brought expanded connectivity options and improved user interaction, adding new data sources, AI providers, and web search capabilities alongside keyboard navigation enhancements.
Highlights
- PWA support enables AnythingLLM to work as an installable web app.
- Keyboard navigation of slash commands now works with arrow keys.
- YouTube video URLs are automatically handled in the uploader and chat.
- SerpApi web search integration added for AI agent queries.
- New data connector for Paperless ngx document management.
New
- PWA (Progressive Web App) support
- SerpApi web search provider
- Z.ai LLM and agent provider
- Paperless ngx data connector
- YouTube URL auto-handling in uploader and chat
- Microsoft Foundry Local for workspace agent provider
Improved
- Keyboard navigation of slash commands now responds to arrow keys on mount
Fixed
- KoboldCPP agent provider max tokens handling
October 2025
1 release: 1.9.0
October brought significant enhancements to AnythingLLM's agent capabilities, adding real-time API access, improved model support, and new ways to handle content. The update expands what agents can do while making the platform more flexible for users.
Highlights
- Agents can now call APIs in real-time without requiring full agent flows
- Agents gained image understanding capabilities for visual content analysis
- All models and providers now support agentic streaming for faster responses
- New real-time file searching skill available by default across all platforms
- Agents can ephemeral scrape or download web resources on demand
New
- Agent real-time API calling without agent flows
- Agent image understanding
- Realtime file searching cross-platform default skill
- Ephemerally scrape or download web resources via agent or uploader
- Microsoft Foundry Local integration
- Agent system prompt passthrough with user context awareness
- Agentic streaming support across all models and providers
Improved
- Default users can now reorder workspaces
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. AnythingLLM: pricing, features and alternatives · All changelogs