Release overview — August 2026

The Curiosity Team

• 2026-08-24

August brought five releases, and one theme runs through them: an assistant that can do more, inside rules you set. Here are the August highlights.

The assistant does more

• Search the workspace in plain language. Describe what you want in your own words; the assistant works out which filters apply and runs a workspace search that respects your access rights. A new LLM Search page controls which data types and filters it may use.
• Deep research runs inside the chat. It plans a question into steps, gathers sourced evidence with the tools it is allowed to use, and writes a report, drawn live as a plan card.
• Skills in a chat, alongside tools and agents, and replies that can render as slide decks.
• Messages sent mid-reply are queued as the next turn instead of being refused, and a Background Tasks panel shows every agent working on the conversation, with its transcript rendered live.

Your models, your clients

• One Models screen lists every provider and model with a per-model toggle. Fifteen API-key providers, including OpenAI, Azure OpenAI, Google, Groq, DeepSeek and OpenRouter, are now configured the same way, and a chat or agent can run on a specific model.
• OpenAI- and Anthropic-compatible APIs expose your configured providers to existing clients and SDKs, authenticated with a bearer token, so those clients never hold the provider credentials.

Rules you set

• Chat policies inspect conversations before the answer, after it, or in batches while the workspace is quiet, and flag them for review or, when run before the answer, refuse them. Two kinds ship: a forbidden keywords list, and a local safety classifier that runs entirely on the workspace server.
• Workspace secrets store a credential once for endpoint and tool code. Values are write-only, never exported, and every read is audit-logged.
• An Access Debugger shows what a user can and cannot see, and explains a verdict on any node as a path through the access groups.
• AWS Cognito and OneLogin single sign-on, plus sign-in behind an AWS load balancer that has already authenticated the user.

Sudo gets a shell

Sudo, the admin assistant, now works on the whole workspace configuration as a filesystem, reading, editing, compiling and diffing it with shell and Python commands that run inside the workspace process. Nothing he writes reaches the workspace until an admin reviews the commit as a side-by-side diff and approves it.

Building on the workspace

• Install NuGet packages from a Packages page, by uploading a .nupkg or pasting a reference snippet. Their types are available to endpoints, tools, tasks and queries immediately, with no redeploy.
• Custom code is checked at compile time for patterns that can outlive its scope, such as Task.Run or async void, and RunTaskAsync is the supported way to run background work.
• curiosity-cli serve --watch rebuilds and reloads a front-end on every change.

Data in and out

• CSV import, rebuilt for large files: the mapping screen previews a 35 MB file in about a tenth of a second, and the import streams in batches that can resume.
• Synthetic exports of any data type keep the real schema and shape but replace the values, so a sample can be shared for testing or support.
• A generated schema reference in Markdown, and facets for yes/no, date and numeric fields.

The little things

• An Active Chats page shows every in-flight chat turn and agent run.
• Access checks are faster: on 60,000 nodes, filtered enumeration went from 63 ms to 32 ms.
• Search results and previews share one row design wherever a node appears, and previews stack as sheets you can step back through.
• Parallel tool calls no longer break AWS Bedrock conversations.

This overview covers builds v26.8.68976 → v26.8.70221. For the full, build-by-build detail see the changelog.