Release overview, October 2026

The Curiosity Team

• 2026-10-20

October has been about running Curiosity Studio yourself and giving assistants more to work with: model providers for self-hosted workspaces, built-in image generation, web search and a browser, OCR that works everywhere out of the box, and server tuning without a restart. Here's the month so far.

More providers, fewer sample tools

• Self-hosted workspaces can pick Curiosity AI as their model provider, during onboarding or from the AI providers list. The administrator receives an activation code by email, and the account credentials stay on the server.
• Two new provider types, OpenAI-Compatible and Anthropic-Compatible, connect to any endpoint that speaks either protocol (vLLM, LM Studio, llama.cpp, Together, Fireworks or a corporate gateway) with an endpoint, a key and a model id. OpenAI providers with a custom host and Azure OpenAI providers move to OpenAI-Compatible on upgrade, with their settings unchanged.
• Inception Labs (Mercury models) joins the AI providers, and the Mistral provider now lists models and runs chats. OpenAI and Anthropic providers need their own API key; use the Curiosity provider for hosted access.
• StructuredAI can send its Choice, Score and Yes/No questions to hosted decision models that return calibrated probabilities: TypeSafe Jev, OpenAI Decisions, Inception Mercury Decide and Cloudflare Workers AI, configured under Manage > AI > Decision Providers. Cloudflare also adds the first extraction model, behind StructuredAI.ExtractAsync.
• Image generation and editing providers (OpenAI, Google Imagen, xAI, DeepInfra, MiniMax and Z.AI) become AI tools assistants and agents can call. Generated images appear in the chat as the picture itself and are saved as the user's file.
• Web search providers (Brave, DeepSeek, SerpApi, Serper, Cloudflare, Tavily, Exa and Perplexity) are offered to assistants as web search tools, the Deep Research template uses whichever are configured, and a provider can be made searchable from workspace search.

The built-in providers replace the old "Create an Image" and "Search the Web" sample tools, which are removed.

OCR on every platform

OCR is now on by default and needs no download. NanoOCR reads and classifies every scanned page or image: printed text keeps that reading, photos go to Florence2, and tables, forms, handwriting and mixed pages go to PaddleOCR-VL once those models are installed. NanoOCR's reading stays as the fallback when a heavier engine is missing or finds nothing.

Sudo and the admin assistants

• A workspace browser gives Sudo and AI-tool sandboxes a headless Chromium to open pages, read them, fill forms and take screenshots, configured under Manage > AI > Browser with the hosts pages may reach. A live view of the page appears in the chat while a reply is browsing, and Sudo uses it to preview the front ends it builds.
• Sudo Graph, docked beside the graph explorer, runs queries and path searches and pushes the results into the view.
• A new Artifacts settings page lists every PowerPoint design system with a preview and an on/off switch, and has Sudo make a new theme from a description, a deck or an existing design.
• New initialization skills walk Sudo through standing up an application on an empty workspace, from use-case discovery to go-live.
• Agents can carry skills, and when an assistant runs a shell command the chat draws it as a terminal window instead of raw JSON.

Building and inspecting

• The admin code editors keep a local edit history next to saved revisions, save with Ctrl+S, and the C# editor gains signature help, an outline, inlay hints, code lenses and quick fixes. Build editors open on the code itself, with settings, tests and previews as docked panes.
• Inspect Chat has a Timeline tab that charts when each turn was thinking, writing, calling tools and waiting, plus tabs listing every model call with its finish reason and every agent run with its status and duration.
• NLP experiments run on one live page and can be stopped, deleted and exported to CSV.
• The loading page is branded with the workspace's logo and name, and administrators can replace it under Front-End > Loading Page.

Tune the server without a restart

• A new Runtime Tuning page (Manage > Configure) adjusts cache intervals, query cache size, memory trimming, simultaneous indexes and linking queue limits live, and sets how the server responds to high memory use. The settings travel with the workspace configuration as config/runtime-tuning.json.
• Memory diagnostics on the Profiling page track memory-management pauses over a window from 1 minute to 24 hours.
• API responses send Cache-Control: private, no-store by default, so proxies and CDNs do not store per-user data.
• Audit events are kept only in the signed audit log, and the new curiosity-cli download-audit-logs command mirrors those files into a local folder, fetching only what is new on each run.

Search and indexing

• EmbeddingGemma 2 (768 dimensions, multilingual) ships in the Docker image and is the default for new AI Search indexes. Other embedding models download only when an administrator asks for them, from Manage > AI > Models.
• Custom code indexes can be marked high priority, and spotters built from a node type pick up new nodes within moments instead of waiting for a rebuild.

For developers

• The Catalyst.Rampart library, which detects personal information and redacts it reversibly on the server, is available to workspace code.
• The upload-front-end-source CLI command commits a local front-end project and queues its build.
• Custom front ends can show an environment badge, set app-wide search defaults with AppSettings.Search, and open, close or handle clicks in the chat citations panel.

The little things

• Up in an empty chat box recalls the messages you sent in the conversation.
• Your own chat messages render as markdown and can be selected.
• A run of tool calls stays in one group when the model thinks between them.
• Reloading a chat that is still running takes seconds instead of minutes.
• Archived chats have one header, no composer, and a phone layout.
• The graph explorer draws every edge between new and existing nodes, and caps "All neighbors" at 100.
• Search filters such as @file and kind: survive reloads and shared links.
• The chat model picker is hidden when only one model is configured.
• Generated avatars show their initials on Linux servers.
• Change History filters by user and by code text.

This overview covers builds v26.10.72149 → v26.10.72602 (October 1 to 20). For the full, build-by-build detail see the changelog.