Oxygen AI Positron for Desktop is an advanced AI-powered solution designed to assist content creators, technical writers, and developers throughout every stage of the writing process. By leveraging the capabilities of AI platforms, AI Positron streamlines content creation and enhances productivity, thus empowering users to craft high-quality documentation and content with greater speed and confidence.

Release Notes 8.4.0

1. Oxygen HTTP MCP Server - The running Oxygen instance can now be exposed as a localhost-only HTTP MCP server from the new Oxygen MCP Server preferences page, which also provides the client configuration. The server is protected with a token and with Origin/Host checks.
2. Review the Changes Made Through the MCP Server - Documents changed by third party AI assistants are listed for review in a dedicated AI Positron toolbar button, from which each of them can be compared with its previous version, kept or rolled back. These changes are kept apart from the ones proposed in the chat.
3. AI Edits Merged into the Open Author Document - When the target document is open in Author mode, AI edits are merged into it as individual changes that respect the current Track Changes state, instead of overwriting the file on disk.
4. The AI Can Ask You Before It Acts - Instead of guessing, the AI now asks you in the chat and waits for an answer whenever your reply would change what it produces. For example, the Translate to actions ask how the translation should be delivered.
5. New models are available:
   • OpenAI:GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, and their Thinking variants. The Thinking variants are not available through the Chat Completions connector.
   • Claude:Sonnet 5 and Opus 5, and their Thinking variants, when using Positron Service or a Claude enterprise connection.
   • VertexAI:Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite, when using Positron Service or the VertexAI enterprise connection.
   • xAI:Grok 4.6, Grok 4.5, and Grok 4.3, when using an xAI enterprise connection.
   • Voice transcription: the speech to text model was updated from GPT-4o Transcribe to GPT Transcribe, for the Positron Service and the OpenAI connections.
6. New Default Model - GPT-5.6 Terra with adaptive thinking is the new default for Positron Service subscriptions and the OpenAI Responses connector. The variant without reasoning is the default for the OpenAI Chat Completions connector.
7. Validate and Check for Completeness Tool - The DITA Agent chat mode can now run a validate and check for completeness on a DITA Map or a DITA OT Project file, with the settings provided by the LLM.
8. Validate With Schematron Tool - All the AI agents can now validate a document either against existing Schematron files or against rules generated on the fly from the requirements you describe, automating compliance with custom business rules. Requires Oxygen 29 or newer.
9. Run Transformation Scenarios from the Chat - The Agent and Dev Agent chat modes can list the transformation scenarios of a document, run one of them, and read the messages it left in the Results panel, so the AI can publish your content and then investigate an unexpected output. A confirmation is always requested before a run.
10. Configurable Maximum Context Window for MS Azure and AWS Bedrock Connections - A new Maximum context window field is available for MS Azure AI and AWS Bedrock enterprise connections, where the context window of a custom deployment cannot be detected automatically. The configured value is used by the context usage indicator.
11. Markdown Response Rendered Visually - Responses in Markdown format are now detected and rendered as HTML, with support for links, tables, lists, headings and inline styling. The new Switch how the response is rendered drop down button changes the visualization mode.
12. View Token Usage Information - The tooltip of the context usage indicator now also shows, for external AI connections, the tokens used by the last response and by the entire conversation, and for Positron Service connections, the plan credits used and the date when they are reset.
13. Automatic Compaction While the AI Is Working - Automatic conversation compaction now also runs between two tool calls, so a turn in which the AI reads a large document no longer fills the context window and fails; the results of the tools that just ran are always kept. It also triggers correctly now on connections whose model is typed in by the administrator, such as an MS Azure AI deployment.
14. Faster Rewrite and Translation Actions - The Correct grammar, Improve readability, Use active voice, Improve DITA structure and Translate to actions now apply targeted edits instead of regenerating the whole content. This is significantly faster on large documents and leaves the markup and indentation they do not touch exactly as they were.
15. Attach Folders in Chat - Folders can now be attached in the chat. Their contents are explored on demand, instead of every file being inlined.
16. Pass secrets to command line tool in a secure fashion - The LLM can now pass environment variables to the command line tool. Oxygen editor variables in the values are expanded automatically, so secret token values are never sent to the LLM.
17. Define sub-categories for AI actions - When the "categoryId" property of an AI action contains multiple values separated by " / ", those values are shown as sub-categories.
18. Recently Used Actions Promoted at the Top - The Recently used section is now presented at the top of the AI actions list.
19. Reason Shown When an Agent or an Action Is Invoked - The chat now shows a one sentence explanation of why a sub-agent or an AI action was invoked, below its name. In chat modes such as Planner, you can tell at a glance what each agent was asked to do.
20. Human-Readable Names Shown for MCP Tools - When an MCP server provides human-readable names for its tools, those names are used in the chat and in the Tools dialog box.
21. MCP Servers Contributed by the Application - Through the new MCPServerConfigurations plugin extension, the host application can contribute HTTP streamable MCP servers of its own, with an authorization header that is recomputed before each request. They are used together with the servers declared in the MCP configuration file.
22. Model Context Protocol (MCP) Improvements - Updated the used MCP SDK to version 2.0.0 to support latest standard changes.
23. Application Context Available to MCP Clients - A new get_oxygen_xml_application_context tool lets third party AI assistants see the context prompt of the project, the documents opened in the editor and in the DITA Maps Manager, the current document and its selection, and the skills available through load_skill.
24. Documents You Visited Included in Metadata - The documents you opened, switched to or edited since your previous message are now listed in the conversation metadata, so the AI knows what you have been looking at.
25. Open Documents Listed in the Application Context - The AI is now told which documents you currently have open in the main editor and in the DITA Maps Manager view, so it can work on them without being told where they are. The lists are kept up to date as you open and close documents.
26. User-Edited Sandbox Files Identified in Metadata - When you edit a file that has pending AI changes in the AI Positron Sandbox, its URL is included in the conversation metadata, so the AI can recognize your edits.
27. Large Images Downscaled Before Being Sent to the AI - Attached images larger than 2000 pixels on either side are now downscaled to that limit, preserving the aspect ratio, to stay within the limits of the vision models and consume fewer tokens. The attached file itself is not modified.
28. Auto-Approve All the Tools of an MCP Server - Entries in the Unsafe tool auto-approval preferences page can now use * as a wildcard in the tool name, so a single entry such as brave-search_* allows every tool of that server to run without confirmation. Entries without a * keep matching exactly.
29. Development Actions Consolidated - The Chat About Code, Explain Code, Suggest Refactoring and Autocomplete AI actions were removed, since better results can now be obtained by chatting with the AI.