> ## Documentation Index > Fetch the complete documentation index at: https://docs.onyx.app/llms.txt > Use this file to discover all available pages before exploring further. # Release Notes > New Features, Improvements, and Bug Fixes ## v3.2.0 Multi-model chat lets you run a prompt across several models at once and pick the best response, a generic OpenAI-compatible provider opens the door to any LLM endpoint, the indexing pipeline gets a significant performance rework, and the Hook system matures with a second hook point and admin status views. Thank you to our contributors for helping make this release possible! ## Key Improvements ### Multi-Model Chat You can now send a single prompt to multiple models in parallel and compare their responses side by side. Pick the response you like best to continue the conversation with that model, or deselect to keep exploring. Multi-model is wired through the chat UI and the Chrome extension. Multi-Modal Chat ### Generic OpenAI-Compatible Provider Onyx now ships a generic "OpenAI Compatible" provider that lets you point at any OpenAI-compatible endpoint — local inference servers, gateway proxies, or self-hosted model runners — without needing a dedicated integration. The Custom LLM provider configuration was also reworked to support this cleanly. ### Hooks: status, logs, and document ingestion The Hook system introduced in v3.1.0 has matured. A new admin page lets you create, edit, and monitor hooks with live status and logs, and a second integration point — document ingestion — joins query processing in the set of hookable lifecycle events. Connection-lost states surface clearly in the UI when a hook becomes unreachable. Hook Extensions Page ## Contributor Highlights * [@alex000kim](https://github.com/alex000kim) added an optional CA certificate update step to the api-server startup in the Helm chart, making it easier to deploy Onyx behind corporate TLS-inspecting proxies! ## Additional Improvements ### [Onyx CLI](https://docs.onyx.app/overview/onyx_anywhere/cli) * `onyx-cli serve` over SSH for working with remote deployments * `onyx install-skill` command * Version compatibility check between CLI and backend * Streaming markdown rendering in CLI output * [Agent-friendly output mode](https://docs.onyx.app/overview/onyx_anywhere/cli#using-as-an-ai-coding-assistant-skill) ### Connectors * Notion connector now extracts people properties and inline table content * Discord connector validates tokens up front ### Chat and search * Files exceeding the token limit can still be uploaded; only indexing is skipped * Citation badges in the embedded chat widget ### LLM providers * Generic OpenAI-Compatible provider type * Custom LLM provider configuration revamp, with first-class fields extracted from the key-value list * Force-delete supported for the last remaining LLM provider ### Auth and Security * Migrated limited-role checks to account-type-based access control * Default Admin and Basic user groups are now seeded automatically ### Performance * Chat sessions query optimized to prevent cascading DB failures under load * Many client-side requests migrated to SWR with a centralized key registry, eliminating duplicate fetches and stabilizing revalidation behavior * Removed unnecessary user-memory eager loads from auth paths * Indexing pipeline streams chunks through a unified iterable-based abstraction with batched processing, improving memory usage and throughput on large corpora ### Deployment and Observability * Grafana dashboard provisioning via Helm, with a dedicated OpenSearch retrieval dashboard * AWS deployment additions: RDS snapshots, EKS CloudWatch logging, RDS CPU / memory / IO alerts, and support for multiple Postgres hosts ### Notable Bug Fixes * Fixed owner-password-only PDFs being incorrectly rejected as protected * Fixed XLSX empty row/column handling * Fixed PowerPoint chart conversion failing during indexing * Fixed paused file connectors not reindexing correctly * Fixed MCP server propagating generic error messages instead of the real cause * Fixed mobile sidebar overlapping content on medium-sized screens, plus several other sidebar and modal responsive issues * Fixed default-provider selection when multiple providers share the same type *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v3.1.0 Voice mode lands, the new Hook system opens up Onyx for extension, the Admin Users and Groups pages get a full rewrite, and we've added two new LLM gateway providers. Thank you to our contributors for helping make this release possible! ## \[Update] Defaults that Changed Two default behaviors changed in v3.1.0 that operators should be aware of before upgrading: * **OpenSearch admin password default removed.** The standard `docker-compose.yml` no longer ships a fallback for `OPENSEARCH_ADMIN_PASSWORD`. If you previously relied on the default, set the password explicitly in your environment before upgrading. ## Key Improvements ### Voice Mode [Onyx now supports voice conversations end-to-end](https://docs.onyx.app/overview/core_features/voice_mode#voice-mode). [Speech-to-text and text-to-speech are configurable from the Admin Panel](https://docs.onyx.app/admins/actions/voice_mode#voice-mode), and chats can be revisited with their voice playback intact. Voice Admin Page ### Hooks Onyx now ships a first-party Hook system for extending behavior at well-defined points in the request lifecycle. [Hooks are managed from a new admin page](https://docs.onyx.app/admins/advanced_configs/hook_extensions#hooks), configured through a connect/manage modal, and currently integrate with the query-processing pipeline, with more hook points planned. ### Redesigned Admin Users and Groups The admin Users page has been rebuilt from the ground up with a stats bar, server-side pagination, role/group/status filters, inline role editing, an invite flow, and edit-group-membership. The Groups page is also new, with list and create/edit pages, inline renaming, shared resources, and token-limit configuration. Users and Requests Admin Page ### LiteLLM and Bifrost gateway providers Two new LLM gateway integrations join the provider list: [**LiteLLM**](https://docs.onyx.app/admins/ai_models/litellm_proxy), which auto-fetches available models from your LiteLLM proxy, and [**Bifrost**](https://docs.onyx.app/admins/ai_models/bifrost#bifrost), configurable from a dedicated provider modal. Both make it easier to centralize model access and routing through your existing infrastructure. ## Contributor Highlights * [@Fizza-Mukhtar](https://github.com/Fizza-Mukhtar) fixed an LLM bug where Onyx was sending `tool_choice` even when no tools were provided! * [@ciaransweet](https://github.com/ciaransweet) corrected the OpenSearch admin password `secretKeyRef` documentation in the Helm values file! * [@benwu408](https://github.com/benwu408) contributed the initial scaffold for a new Canvas connector! ## Additional Improvements ### Connectors * SharePoint connector gained deduplication, filters/denylists, non-public shareable links, and better error logging * Notion connector upgraded to the latest API version * Confluence client retries on transient failures * Jira large-batch JSON decode fix * Google Drive: rate-limit retry and fallback access path for externally owned items * Discord connector now properly cleans up async resources ### Agents and Chat * Tool call arguments now stream live, with a delta-based jsonriver implementation * Custom tools have improved error handling and a refreshed timeline UI * Configurable file upload size limit, with a frontend precheck for oversized uploads. This can be configured in the Admin Panel, under Advanced Options for Chat Preferences. File Attachment Size Limit * More file types can be attached to chats, and the file preview supports significantly more formats. File types which are not explicitly supported/recognized are parsed using a text-detection heuristic. * Slackbot now resolves channel references and can filter search results by channel * Search state machine consolidated for more predictable behavior ### Auth and Security * PKCE support for OIDC ### LLM providers * LiteLLM and Bifrost gateways * All provider types can now be deleted (previously some were soft-locked) * Editing an LLM provider correctly uses the global default model * Surface masked OpenAI quota failures so users see a real error * Filter embedding models out of LiteLLM model lists ### Performance * Stop deep-copying during indexing * Plaintext file results are now cached in chat * OpenSearch tuning continues with configurable shards/replicas, configurable Vespa migration page size, configurable hybrid subquery hits, vectors excluded from hybrid/random search payloads, and Prometheus metrics for retrieval ### Deployment * Windows installer * Terraform module for OpenSearch on AWS * Configurable `api_server` resource limits in Docker * Reusable Prometheus metrics server for Celery workers, with queue depth, connector health, task lifecycle, and per-connector metrics ### Notable Bug Fixes * Fixed CSV inline display responsiveness and comma-literal parsing * Fixed memories immediately losing focus on click * Fixed numerous modal centering, popover, sidebar, and Safari-specific UI issues *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v3.0.0 The Vespa to OpenSearch migration is here, SCIM 2.0 lands for Enterprise, search and chat are unified into a single experience, Code Interpreter ships by default, and Onyx Lite and the Onyx CLI debut. v3.0.0 is the largest release we've ever cut. Thank you to our contributors for helping make this release possible! ## \[Update] OpenSearch Migration v3.0.0 ships the full Vespa to OpenSearch migration toolchain. OpenSearch now runs alongside Vespa in the standard `docker-compose.yml` and Helm chart, and admins get a new Migration tab in the Admin Panel for kicking off and monitoring the migration with live progress. Vespa continues to serve all queries throughout the migration by default. Once the migration completes, admins can switch Onyx to read from OpenSearch. Setup details and guidance live in the [OpenSearch deployment guide](https://docs.onyx.app/deployment/local/opensearch), and the full migration walkthrough is in the [Vespa to OpenSearch migration guide](https://docs.onyx.app/admins/advanced_configs/opensearch_document_index_migration). ## Upgrade Path to v4.0.0 v3.0.0 starts the final stretch of the Vespa to OpenSearch migration. **Every deployment must run a v3.x.x release long enough for the migration to complete before upgrading to v4.x.x, where Vespa will be fully removed.** If you skip v3 entirely, you will lose all of your indexed data and have to re-index from scratch — including re-uploading any Project files. ## Key Improvements ### SCIM 2.0 User and Group Provisioning Onyx Enterprise now ships a complete SCIM 2.0 implementation. Identity providers like Okta and Entra ID can create, update, and deactivate users and groups against Onyx via the standard `/scim/v2/*` endpoints with bearer-token auth. Admins can mint and rotate SCIM tokens from a new section in the Admin Panel, and we've passed Okta's Runscope spec test for OIN submission. [More details can be found in our documentation.](https://docs.onyx.app/deployment/authentication/scim#scim) ### Unified Search and Chat Search and chat are now driven from the same input bar, with smooth animated transitions between modes. Whether you're asking a question of an agent or running a quick keyword search across your knowledge base, you start in the same place. ### Code Interpreter, by default Code Interpreter is now part of the default deployment for both Docker and Helm — no extra setup required. Sessions stream output as they run, accept file inputs, embed generated images directly into chat, and ship with rich preview modals for CSV, PDF, and DOCX outputs. The Python tool is also enabled by default on the default agent. [More information can be found here.](https://docs.onyx.app/admins/actions/code_interpreter) ### Onyx CLI A new command-line tool for working with Onyx, distributed as a PyPI package and a Docker image. Look for `onyx-cli` to grow over the next few releases. [Get started here.](https://docs.onyx.app/overview/onyx_anywhere/cli#cli) ### Onyx Lite A new "lite" deployment mode for installations that don't need a vector database. Onyx Lite runs without the indexing stack and is suitable for chat-only and tooling-focused use cases. ## Contributor Highlights * [@victoriamreese](https://github.com/victoriamreese) added a configurable MCP host setting and continued to harden Helm and KEDA scaling for self-hosted deployments! ## Additional Improvements ### Connectors * SharePoint connector received a major scalability rework, including delta sync, more resilient credential refresh, better permission sync, and substantially better throughput on large multi-site tenants * Slack connector can now optionally include bot messages during indexing * Slack bot responses convert markdown tables to a Slack-friendly format and sanitize HTML and broken citation links * Confluence and Google Drive connectors are now backed by hierarchy nodes for more accurate pruning ### Agents and Chat * Featured agents with labels can be highlighted in the agent picker without being public * New chat sharing UI with cleaner permissioning * Refreshed admin Chat Preferences page * Infinite scroll in the chat session sidebar * Configurable MCP host setting * Tab reading from the Chrome extension side panel * Tooltip and timeline polish, including better tool status messages and a fix for double-collapse in agent timelines * Throughout the codebase and UI, "assistant" identifiers have been renamed to "agent" for consistency, and "LLM Models" is now "Language Models" * LM Studio is available as a new LLM provider ### Auth and Licensing * Enterprise license enforcement is now on by default; deployments without a valid license fall back to a gated mode * Slack users are excluded from seat counts; deactivated Slack users can be reactivated when seats are available * New invite-whitelist setting with a dedicated "account blocked" page for users outside the whitelist ### Deployment and Helm * Production Prometheus instrumentation for the API server, including SQLAlchemy connection pool and per-tenant metrics * Code Interpreter is included in the default Docker and Helm deployments * New Postgres-backed cache backend abstraction for caches that previously required Redis ### Notable Bug Fixes * Fixed Bedrock tool calling configuration and tool content handling without `toolConfig` * Fixed Vertex AI Claude Opus 4.6 model selection and sorted Vertex model names * Fixed Ollama context-window passthrough, content skipping, and DeepSeek handling * Fixed desktop app link clicking behavior * Fixed basic users being unable to share agents * Fixed `/llm/provider` returning a filtered set of providers * Fixed model configuration not populating during sync * Fixed metadata file handling for larger ZIPs * Fixed SharePoint credential refresh and `driveItem` permission sync * Fixed `open_url` on non-normalized URLs and SearXNG web search * Fixed search tool being enabled when no knowledge sources were selected * Fixed Google connectors redirecting to an auth error page after disconnect * Fixed Gong connector to respect the `Retry-After` header * Multiple UI fixes across modals, dropdowns, sidebar, whitelabel logo, code blocks, popovers, and the projects modal *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.12.0 A redesigned agent message experience, Brave web search, and the OpenSearch migration enters its final stretch. Thank you to our contributors for helping make this release possible! ## \[Update] OpenSearch Migration We've begun migrating Onyx Cloud to OpenSearch! So far, the migration has progressed smoothly and we expect it to be completed by the end of the week. The v3.0.0 beta tag should be promoted very soon. As a reminder, everyone must upgrade to a v3.x.x version before moving on to v4.x.x, where Vespa will be fully removed. ## Key Improvements ### Redesigned Agent Message Timeline Agent messages have been refreshed with a new timeline experience. When an agent processes your request, you now see a step-by-step, expandable timeline showing exactly what the agent is doing. Agent Message Timeline ### Brave Web Search Brave is now available as a web search provider! Configure it via the Web Search page in the Admin Panel with your Brave API key. Brave Web Search ### Chat History Compression Long conversations now automatically summarize older messages to keep context windows efficient. This reduces token consumption for extended chats while preserving the key information from earlier in the conversation. ## Contributor Highlights * [@victoriamreese](https://github.com/victoriamreese) enabled optional host settings and resolved pod label duplication in Helm charts! ## Additional Improvements ### Chat and Project Search You can now search across your chat history and projects directly from the sidebar. The new command menu supports keyboard navigation, infinite scroll, and fast filtering to quickly find past conversations. ### Helm and Infrastructure * Added health probes for all containers in Helm charts * Fixed Ingress template generation * Fixed PostgreSQL operator labeling * Improved IRSA support by passing the correct AWS region ### Notable Bug Fixes * Fixed Bedrock tool calling configuration * Fixed Ollama model selection in the UI * Fixed desktop app link clicking behavior * Fixed LLM provider configuration not syncing model flow data * Fixed search tool being enabled when no knowledge sources were selected * Fixed private models appearing in Agent creation * Fixed LLM read timeout handling * Fixed Asana connector workspace/team ID mismatch * Fixed Salesforce connector cleanup logic * Multiple UI fixes across modals, dropdowns, agent creation, and sidebar rendering *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.11.0 Custom chat backgrounds, a refreshed user settings page, the desktop app matures, and continued progress on the OpenSearch migration. Thank you to our contributors for helping make this release possible! ## \[Update] OpenSearch Migration Our migration from Vespa to OpenSearch (announced in [v2.8.0 notes](#\[announcement]-vector-database-migration)) continues to make good progress. We've successfully migrated our internal Onyx instance to OpenSearch! The migration processed \~200k chunks/hour on a Helm-based cluster with 128GB Vespa (r6i.4xlarge nodes), 8GB OpenSearch, and 8GB migration worker memory limits. Remaining work includes polishing the migration flow, improving performance, and final QA. We expect v3.0.0, the stable release of the OpenSearch migration, to be released in the final week of February. ## Key Improvements ### Custom Chat Backgrounds You can now personalize your chat experience with custom backgrounds! Choose from a selection of options in the new Appearance section of the User Settings page. Custom Dark Chat Backgrounds Custom Light Chat Backgrounds ### Refreshed User Settings Page The user settings page has been completely redesigned with a cleaner layout and better organization, making it easier to manage your profile, preferences, and backgrounds all in one place. Refreshed User Settings Page ### Desktop App Updates The Onyx Desktop App received several important updates in this release: * **Domain configuration** — You can now configure which Onyx instance the desktop app connects to directly from the app * **Proper macOS code signing** — The Mac app is now properly signed, so you won't see Gatekeeper warnings on install * **Light/dark mode** — The desktop app now correctly reflects your theme preference ## Contributor Highlights * [@ciaransweet](https://github.com/ciaransweet) clarified the send-chat-message API docs to document the `ChatFullResponse` response model! ## Additional Improvements ### User MCP Actions Users can now attach MCP actions to their agents directly. This is an early step toward giving all users the ability to extend their agents with custom tools and integrations. ### Allow No System Prompt Agents can now be configured with no system prompt, giving you full control over the model's behavior for use cases that require an empty system context. ### Prompt Shortcuts Prompt shortcuts now enforce uniqueness constraints and received several usability enhancements. ### Notable Bug Fixes * Fixed Azure OpenAI tool calling and LiteLLM Azure model streaming issues * Fixed streaming text appearing in bursts after citations * Fixed citation sidebar not appearing for web\_search-only assistants * Fixed Ollama model selection in the UI * Fixed Asana connector workspace/team ID mismatch * Fixed Salesforce connector cleanup logic * Improved Slack query expansion to filter garbage from person names * Added Redis TTLs to prevent potential memory issues * Multiple UI fixes across modals, dropdowns, LLM provider configuration, and agent creation *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.10.0 An update on the OpenSearch migration, Onyx in your Discord server, expanded image generation support, and multilingual Deep Research. Thank you to our contributors for helping make this release possible! ## \[Update] Vector Database Migration A small change in plans for our migration from Vespa to OpenSearch (announced in [v2.8.0 notes](#\[announcement]-vector-database-migration)). Instead of v2.11.0, the migration version will be **v3.0.0**. Vespa will be deprecated in **v4.0.0**. Our intention is that everyone must upgrade to a **v3.x.x** version before moving on to **v4.x.x**. Additionally, we do not anticipate many minor versions of v3 before moving to v4. If you skip a v3 version, you will lose all of your indexed data! This means re-indexing all of your data and re-uploading your files in Projects. Everything else remains the same. When you upgrade to a v3 version, admins will be prompted to start the migration process and Vespa will continue to power your search until v4.0.0. To accomodate this change, our timeline has been pushed by 2 weeks: * **v3.0.0 Beta** will be released in the second week of February * **v3.0.0 Stable** is slated for the third week of February We will continue to share details, migration guides, and any changes to the plan moving forward! ## Key Improvements ### Onyx Discord Bot You can now chat with your agents and search your knowledge base directly from Discord! OnyxBot is available in our [Discord server](https://discord.gg/4NA5SbzrWb). Just tag `@onyxbot` in your message to start a conversation. You can install our [official OnyxBot Discord app](https://discord.com/oauth2/authorize?client_id=1460722463431327804) or deploy your own Discord bot with your self-hosted Onyx instance. Check out the [Discord Bot documentation](/admins/miscellaneous/discord_bot) to get started! Onyx Discord Bot ### Nano Banana Image Generation You can now configure Nano Banana as an image generation provider. Add Vertex AI credentials in the **Image Generation** page of the Admin Panel to use Gemini 2.5 Flash Image (Nano Banana) and Gemini 3 Pro Image (Nano Banana Pro). Nano Banana Image Generation ### Multilingual Deep Research Our Deep Research mode was previously biased toward English responses. We've tuned it to work and respond in the language you ask your question in, making Deep Research more accessible for international users. Multilingual Deep Research ## Additional Improvements ### Slack Connector Restored Slack has rolled back changes to their API that previously prevented indexing. For the time being, we recommend using the indexing connector if possible, as the search quality is significantly better. If you don't see the indexing connector, make sure you've removed all Slack Federated connectors first. Note that if you switch from the federated connector, you will lose access to searching group DMs and direct messages. Read more about the [Slack connector](/admins/connectors/official/slack/slack_indexed) and [Slack Federated connector](/admins/connectors/official/slack/slack_federated) in our documentation. ### Agent Access Controls Restored User and group access configuration was briefly removed from the Agent configuration page during refactoring. This interface has been restored! Pre-existing agents that had access controls configured continued to respect those settings during this period. ### Empty Packets Fix If you've encountered an error like "message failed due to empty packet," we've implemented a fix and you should no longer see this issue. This fix was particularly targeted at models hosted with vLLM. ### Chat Naming Fix for Long Messages Previously, if your very first interaction was too long, the chat session could be renamed to an error message about exceeding context limits. We've adjusted this behavior and chat naming now works correctly regardless of message length. *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.9.0 Two brand new ways to access Onyx and a bumper crop of bug fixes! Thank you to our contributors for helping make this release possible. ## Key Improvements ### Onyx Chrome Extension Chrome(-ium) users can now have Onyx in every tab! The Onyx Chrome Extension brings your Agents, chat, and search to where you are. Features include: * **Quick-access address bar** - Type your question directly in the address bar * **New-tab page** - Start every browsing session with Onyx * **Contextual insights** - Highlight any text on a webpage and birng it into your chat * **Always-on sidebar** - Open Onyx as a sidebar on any webpage Check out the [Chrome Extension](/overview/onyx_anywhere/chrome_extension) page to get started! Onyx Chrome Extension ### Onyx Widget You can now embed an Onyx chatbot into any website! Easily add an AI agent or knowledge search to your website with a single line of code. The widget abstracts our chat API and can be configured to use our Cloud or your self-hosted instance of Onyx. Requests from the widget must use an Onyx API key. We recommend using a Limited API key and monitoring usage to avoid unexpected costs. Read more about [deploying the widget here](/deployment/configuration/website_widget) or check out the [widget README](https://github.com/onyx-dot-app/onyx/blob/main/widget/README.md)! Onyx Widget ## Contributor Highlights * [@coreyauger](https://github.com/coreyauger) updated some out-of-date links in our READMEs! ## Additional Improvements ### In-app Notifications System We've added a new notifications system to Onyx! You'll see in-app release notifications and feature announcements. We have no intentions to spam 😅! Depending on feedback, we can add or remove notification frequency. \-- Editor's note: Apologies for the belated release of the v2.9 notes. It's been a busy start to 2026 and we'll try to resume our regular schedule of releases. As always, our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.8.0 Big cost savings, better cost tracking (and observability), and a heads up for v2.11! Thank you to our contributors for helping make this release possible! ## \[Announcement] Vector Database Migration We've recently made the decision to migrate the Onyx vector database from Vespa to OpenSearch. OpenSearch is a FOSS fork of ElasticSearch, and will be deployed similarly to the existing `index` container. We expect OpenSearch to be easier to develop with, provide equal if not better search performance, and, importantly, have a smaller memory footprint. The official migration from Vespa to OpenSearch will begin in February, 2026, with the **v2.11.0** release. In v2.11.0, admins will be prompted to start a migration process, which will copy over all indexed data from Vespa to OpenSearch. Onyx will remain fully functional with Vespa in this version. Following v2.11.0, we plan to release v3.0.0, which will use OpenSearch exclusively. This timeline is not set in stone. We will publish a detailed migration guide in the coming weeks and communicate any changes! Onyx X OpenSearch ## Key Improvements ### Prompt Caching Cost savings We've added prompt caching support for providers that require explicit caching parameters. This could reduce total LLM costs by up to 50-90%, depending on your usage patterns and your models/providers. Anthropic, Gemini, and Bedrock users should see the largest savings. You may not see any cost reduction if you are using OpenAI or Azure OpenAI, since they implicitly cache prompts for models newer than GPT-4o. ### Cost Tracking We've added LLM cost tracking and time to first token metrics for tracing providers like Braintrust. See our [deployment configuration docs](/deployment/configuration/configuration#observability) for how to set up observability for your deployment. ## Additional Improvements ### Azure Image Generation Models Added support for Azure image generation models for GPT Image 1.5 and GPT Image 1. ### Various Bug Fixes and Improvements * Agents can now "open" PDF files to read their contents in the internal search flow * Fixed a web connector bug where certain pages served as raw bytes were not being parsed as text * Fixed Mistral model tool calling support * Multiple fixes to frontend styling, component rendering, markdown formatting, and performance *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.7.0 A brand new Agent creation page, LLM configuration forms saw some love, and a new Image Generation page! Thank you to our contributors for helping make this release possible! v2.7.0 introduces [a new API](/developers/guides/chat_new_guide) for sending messages to Onyx. The old `/chat/send-message` and `/chat/send-message-simple-api` endpoints will be removed by Feb 1st, 2026. ## Key Improvements ### Agent Creation The Agent creation page received a makeover! The form is now simpler, more intuitive, and matches the (sleek and beautiful) design of the rest of the Onyx interface. Agent Creation Page ### Image Generation Configuration Gone are the days of setting environment variables for image generation. There's a new page for configuring image generation models in the Admin Panel. You can re-use your existing OpenAI API key or setup new OpenAI or Azure providers. Nano Banana support is coming soon! If you've previously set up an image generation model via environment variables, you will need to create a new configuration via the Admin Panel to continue generating images. Image Generation Configuration Page ### New LLM Configuration Forms The LLM configuration form also received a refresh in this release. If you're using OpenAI, Anthropic, Vertex/Gemini, or OpenRouter, you can now set the providers to "auto-update" and display the latest, most recommended models. The auto-update feature requires an internet connection to GitHub to fetch our recommended models list. LLM Configuration Forms ## Additional Improvements ### MCP Server Deployment The Onyx MCP Server is now disabled by default in the `docker-compose.yml` file. You will no longer see the container start and stop (when disabled) in Docker. To deploy the MCP Server, you will need to set `MCP_SERVER_ENABLED=true` in your environment and uncomment the MCP Server section in your `docker-compose.yml` file. ### Model Server Changes Multiple custom models were removed from the model server image. As a result, Onyx's total disk and memory footprint was reduced! ### New Cloud Usage Limits We've recently seen a wave of abuse of our Cloud service's free LLM access. Moving forward, if you use our provided LLM keys, you'll see a weekly limit on LLM usage cost. If you run into this limit, feel free to reach out to us at [hello@onyx.app](mailto:hello@onyx.app), and we'll be happy to help! These changes only apply if you use Onyx-provided LLM keys on [Onyx Cloud](https://cloud.onyx.app?utm_source=docs\&utm_content=changelog). *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.6.0 Welcome to the Onyx v2.6.0 release notes! This is one of our **largest releases to date**, and it includes a completely rebuilt chat experience, the Onyx Desktop App and MCP Server, and smarter model management that brings you the latest models from your providers without additional effort. Thank you to our contributors for helping make this release possible! **Known issue in v2.6.0:** During an embedding model swap, you will not be able to search your knowledge base until the swap is complete. This functionality will be restored in the next major release. ## Key Improvements ### Chat Experience Revamped In v2.0.0, we overhauled the visual design of Onyx. In this release, we've dug into the chat internals to make it more powerful, work with more models, and ultimately deliver the best LLM experience possible. To do this, we've built custom architecture for managing conversations and context, introduced parallel tool calling, and added the ability to invoke agents as tools in the chat. The result is an AI chat experience where your agent can find information from your knowledge base and the internet, keep only the most relevant information in context, call tools precisely when needed, and answer your questions quickly and accurately. Our Deep Research mode makes use of all of these new features. Try it out and let us know what you think! Deep Research with new chat capabilities ### Dynamic Model management No more redeploying to access new models. Onyx now fetches model information dynamically with richer metadata, which means: * New models are automatically retrieved and work out of the box, including vision-capable models * Deprecated models are automatically hidden from your interface * When max tokens can't be detected, we now fall back to 32k instead of the previous 4k limit Note that admins will still need to make new models visible for end users via the LLM configuration page. Dynamic model management ### Onyx MCP Server Bring Onyx directly into your development workflow with our new MCP server! Connect Onyx to Claude Code, your terminal, or any MCP-compatible client. Authenticate with your PAT or API key and access your document index and web search capabilities without leaving your tools. Upcoming features include OAuth authentication and uploading and managing files. Onyx MCP Server ### Onyx Desktop App (Preview) We're developing a native desktop application! The Onyx desktop app will be available on Windows, macOS, and Linux. If you would like to try it out, you can download the app from the [v2.6.0 release page](https://github.com/onyx-dot-app/onyx/releases/tag/v2.6.0#:~:text=21%20other%20contributors-,Assets,-10). Through the desktop app, you can access Onyx Cloud or your self-hosted instance. Onyx Desktop App The Onyx desktop app is currently published as a development/preview build. Please be aware that there may be bugs and issues. Documentation will be available when the app is officially published. ## Contributor Highlights * [@Django149](https://github.com/Django149) overhauled the file connector. You can now directly upload and remove files from the connector without creating a new one! * [@brano-rohlik](https://github.com/brano-rohlik) added Gemini 3 Pro and Claude Opus 4.5 support! * [@ethanwater](https://github.com/ethanwater) added a new connector for Coda! * [@rexjohannes](https://github.com/rexjohannes) added a Drupal wiki connector! * [@emerzon](https://github.com/emerzon) went on a tear this release! He: * Hardened markdown link processing to prevent potential security vulnerabilities * Removed duplicate DB operations to speed up chat deletion * Fixed a connector state bug, resulting in fewer connector failures * Added LiteLLM user and session tracking for budget control and observability * [@kiwamizamurai](https://github.com/kiwamizamurai) fixed an issue with the DISABLE\_MODEL\_SERVER environment variable not being respected! * [@vinitngr](https://github.com/vinitngr) cleaned up the install script to use a consistent root directory! ## Additional Improvements ### SearXNG Web Search Support We've added SearXNG as a web search provider, giving you another self-hostable choice for powering your agents' internet access. Through SearXNG, you can configure multiple search engines and keep your search results private and secure. ### Improved File Handling in Chat Files uploaded to chat are now processed more reliably, with the ability to preview them directly in the interface. Improved File Handling in Chat ### MCP Passthrough Authentication MCP connections now support passthrough authentication for seamless integration with your existing auth systems. ### Embedding Model Configuration You can now update API keys for embedding models without recreating the configuration. ### Bug Fixes * Multiple fixes to Confluence, Jira, and web connectors * Improved document processing stability, reducing likelihood of OOM errors when using API-based embedding models *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.5.0 Welcome to the Onyx v2.5.0 release notes! This update introduces code execution, new web search providers, and various improvements across the platform. Thank you to our contributors for helping make this release possible! ## Key Improvements ### Web Search Enhancements We've officially released support for new web search providers including Google PSE, Serper, Firecrawl, and a built-in Onyx web scraper! Configure these providers via the Web Search page in the Admin Panel and give your agents the ability to search the internet for real-time information and current events. New APIs for web search are also available for developers. Detailed setup and API reference docs will be coming soon! Web Search Providers ### Claude Opus 4.5 Support Onyx now supports Anthropic's newest, most powerful model: Claude Opus 4.5! Configure Claude Opus 4.5 via the LLM Configuration page in the Admin Panel. Claude Opus 4.5 ### Code Execution (Alpha) Onyx's code execution is now in alpha! This feature allows your agents to write and execute Python code to solve complex problems, perform data analysis, and generate visualizations. Code Execution Alpha We built code execution, in-house, to work with any model provider and LLM. It's a secure, FastAPI-based service that uses isolated Docker containers with strict resource limits, timeout controls, and file handling capabilities for security and reliability. Code execution is designed to work with any AI agent, not just Onyx. It's available as a standalone, fully open source project: [onyx-dot-app/code-interpreter](https://github.com/onyx-dot-app/code-interpreter?utm_source=docs\&utm_content=changelog). While Python is the first supported language, all common languages will be supported in the future. Code execution is not yet ready for production use. If you enable the feature flag, you may encounter bugs! Please bear with us as we continue to develop and improve this feature. ## Contributor Highlights * [@sashank-rayapudi-ai](https://github.com/sashank-rayapudi-ai) added a new connector for Testrail! * [@mristau-alltrails](https://github.com/mristau-alltrails) identified and resolved a vulnerability in our nginx configuration! ## Additional Improvements ### Default Assistant MCP Tools The Default Assistant can now have MCP tools attached, making it easier to extend the default experience with custom actions and integrations. ### Document Search APIs in Community Edition Document search APIs have been moved from Enterprise Edition to Community Edition, giving all users access to programmatic document search capabilities. ### Auto-Pause for Failing Connectors Connectors that continuously fail will now automatically be paused. This helps relieve pressure on the background, API, and indexing containers, improving overall system stability. ### Kubernetes Affinity and Tolerations We've added affinity and tolerations to the Helm deployment templates, allowing you to customize where pods are scheduled in your Kubernetes environment. *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.4.0 Welcome to the Onyx v2.4.0 release notes! This release introduces powerful new controls for the Slack Federated connector, fine-grained LLM access controls, and various other improvements and bug fixes across the platform. Thank you to our contributors for helping make this release possible! ## Key Improvements ### Advanced Filtering for Slack Federated Connector The Slack Federated connector now supports comprehensive, entity-based filtering, which gives you precise control over the data available to Onyx users. Configure filters such as: * Channel selection with glob patterns * Message type filtering (1:1 DMs, group DMs, private channels) * Date extraction If you have an existing Slack Federated Connector, you will need to update the app's scopes to use the expanded search functionality (e.g. searching group DMs). Additionally, each user will need to revoke and re-authorize their OAuth access to the app for the new scopes to take effect. See the [Slack Federated Connector](/admins/connectors/official/slack/slack_federated/) app manifest documentation for more details. Slack connector filtering options ### LLM Access Controls Admins can now configure fine-grained access controls for LLM providers. This feature allows you to restrict which users or groups can access specific language models. Use access controls to better manage and monitor generative AI usage across your organization. Read more about LLM access controls in our [admin docs](/admins/advanced_configs/llm_access_controls/). LLM access controls ### Disable Default Assistant You now have the flexibility to disable the default assistant in your Onyx workspace. This is particularly useful for organizations with many custom assistants, each with specific configurations, tools, and instructions. To disable the Default Assistant, go to the **Workspace Settings** page and toggle the **Disable Default Assistant** switch. ## Contributor Highlights * [@sktbcpraha](https://github.com/sktbcpraha) fixed an issue where the Default Assistant was overriding legacy custom agent prompts! * [@alex000kim](https://github.com/alex000kim) fixed a tricky issue with special characters in the Teams connector! ## Additional Improvements ### Share Chat Button We've made sharing chats easier with a dedicated share button in the top right of the chat interface. ### Paginated Document Search API The document search API now supports pagination for more efficient retrieval of large result sets. ### Projects Token Limits We've increased the per-file token limit for Projects to 100k. Additionally, you can configure this limit with the `SKIP_USERFILE_THRESHOLD` environment variable. ### SSO User Provisioning Improvements Previously, inviting users to join your Onyx workspace would prevent non-invited users from automatically registering. Now, SSO-enabled workspaces can have both invited users and JIT-provisioned users! *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.3.0 Welcome to the Onyx v2.3.0 release notes! This release brings performance improvements that make Onyx feel significantly faster and more responsive, personal access tokens, expanded LLM provider support, and more. Thank you to our contributors for helping make this release possible! ## Key Improvements ### Personal Access Tokens Onyx now supports Personal Access Tokens for API authentication. This feature allows all users to have a secure, programmatic way to interact with Onyx via the API, making it easier to integrate Onyx into your workflows and automation tools. Personal Access Tokens ### Massive Frontend Performance Improvements We've made extensive performance optimizations across the frontend that deliver a noticeably faster and more responsive experience. * Typing in the input bar is now lag-free, even during heavy operations. * UI updates are significantly faster with improved reactive performance, making interactions feel instant and smooth throughout the application. * First content load is also much faster with optimizations in data loading. Frontend Performance Improvements ### Expanded LLM Provider Support We've added support for OpenAI image generation models via Azure, bringing more flexibility to your image generation capabilities. Additionally, Onyx now supports AWS GovCloud regions for Bedrock. ## Additional Improvements ### White Labeling Fixes We've resolved several issues with white labeling to ensure your custom branding displays correctly across all areas of the application. ### More Resilient Gmail Connector The Gmail connector now uses checkpointing for more resilient indexing. This makes the connector more reliable and significantly cheaper to run, especially when re-embedding your entire Gmail history. ### Air-Gapped Deployment Support We've fixed air-gapped deployments by packaging the tokenizer directly with the model server image. This ensures Onyx can run in completely isolated environments without requiring external dependencies. ### Optimized LLM Interactions We've simplified both system prompts and tool prompts, resulting in lower token consumption per interaction and more efficient tool calling. This reduces your LLM costs while maintaining the same quality of responses. *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.2.0 Welcome to the Onyx v2.2.0 release notes! This release brings major improvements to the agent framework with faster responses and more consistent outputs, a completely redesigned onboarding experience, and significant UI performance enhancements. Thank you to our contributors for helping make this release possible! ## Key Improvements ### Complete Agent Framework Revamp We've completely rebuilt the agent framework that powers Onyx conversations. The new implementation moves away from LangGraph to the Agents SDK, resulting in a dramatically simplified flow that delivers more consistent outputs, faster response times, and better tool calling accuracy. This architectural change also reduces per-message token consumption, making your conversations more efficient and cost-effective. ### New Login UI and Onboarding Flow First-time users now experience a streamlined onboarding journey that guides them through personalizing their workspace, setting up their LLMs, and configuring actions like Web Search and Image Generation. Onboarding flow ### UI Performance and Polish We've made extensive improvements across the UI for performance and visual consistency. For example, the Assistants navigation page has been completely refreshed with a cleaner layout and improved organization. Agents navigation ### LLM Usage Observability Onyx now supports integration with [Braintrust](https://usebraintrust.com) or [LangFuse](https://www.langfuse.com) for comprehensive LLM observability. Monitor your costs, latency, token usage, and more to gain insights into your AI operations. Read more about setting up observability in our [deployment configuration docs](/deployment/configuration/configuration/). LLM usage observability ## Contributor Highlights * [@dafeliton](https://github.com/dafeliton) fixed an elusive bug with the web connector not parsing PDF URLs correctly and crashing! ## Additional Improvements ### Confluence OAuth and Jira Fixes We've resolved issues with Confluence OAuth authentication and updated Jira integration to address deprecated API endpoints, ensuring reliable connectivity with your Atlassian tools. ### Enhanced Provider Support Onyx now has improved support for LLM providers like Ollama and OpenRouter, giving you more flexibility in choosing the models that work best for your use case. *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.1.0 Welcome to the Onyx v2.1.0 release notes! This is a minor update with new support for OAuth Actions, improvements to the developer experience, and a new FOSS repo. Thank you to our contributors for helping make this release possible! ## Key Improvements ### OAuth Actions Onyx now supports Actions that require OAuth authentication. With this new feature, Onyx can now invoke many more actions on your behalf. Admins can configure OAuth Actions via the OpenAPI tab of the Actions page. OAuth Actions configuration ### Improved Code Rendering Code blocks in chat now render much more consistently across all languages with proper syntax highlighting. This makes it easier to read and understand code generated by Onyx. Code rendering ### Fully Open Source (FOSS) Repository We're excited to share a completely FOSS version of Onyx! All code is MIT licensed and Enterprise Edition code is removed. This FOSS repo should make it easier for developers to contribute, audit, and deploy Onyx without licensing concerns. We will continue to develop on the main `onyx-dot-app/onyx` repository, and `onyx-foss` will automatically sync new changes. Check out the repo at this link: [onyx-foss](https://github.com/onyx-dot-app/onyx-foss?utm_source=docs\&utm_content=changelog). ## Contributor Highlights * [@camro](https://github.com/camro) fixed a tricky issue where the app was not redirecting to shared chats on login! ## Additional Improvements ### Name Personalization You can now set your name in the bottom left user icon! Just set your name in the User Settings modal. Your name and role are used to better personalize your conversations in Onyx. Name personalization ### Theme Preference Persistence Your light mode/dark mode preference is now saved across browsers and sessions, even after cache clears. Set it once and forget about it! ### Image Size Reduction We've fixed a bug that was doubling the size of the model server Docker image. This reduces disk space requirements significantly for deployments (\~26 GB needed for a fresh deployment). *** Our [GitHub Releases](https://github.com/onyx-dot-app/onyx/releases?utm_source=docs\&utm_content=changelog) page has the full list of changes and bug fixes for each release. ## v2.0.0 v2.0.0 has breaking changes. You may want to test the upgrade in a staging environment before deploying to production. Today, we're excited to announce the release of Onyx v2.0.0! This release brings a brand new UI, Projects, and many other improvements. Thank you to our contributors for helping make this release possible! ## Key Improvements ### UI Refresh Onyx has a brand new UI with updated colors, components, and subtle changes to the layout. Check out this release video showcasing the new UI!