What's new in WorkspaceGPT

Calls are detected on Windows too, and notes are kept on your machine. Call notes also stop a stale recording and keep going when transcription falls behind.

A ticket's unread attachments are listed, and WorkspaceGPT reads one only after you agree.

Create PR writes the commit message from the actual diff, and a slow push gets ten minutes instead of timing out. Repositories are looked up with each of your GitHub accounts until one can read them.

When a sync, model call, code host or model download fails on an untrusted TLS certificate, you now get a message that says so and what to do about it.

Dictate into the composer on Windows with an on-device recognizer. Unsent drafts are saved reliably, and a failed save is reported instead of silently lost.

Models that Copilot serves only through its responses endpoint now run through the desktop app.

When a call starts, WorkspaceGPT asks whether to take notes, or does so automatically if you turn that on. On macOS the call is transcribed on-device with Whisper. Notes sit in the sidebar and on your work home, and the agent can use them.

Reports now end with the steps to verify the change. Python projects get real verification, and the counts of what changed are honest.

Pressing Run on a shell block shows its output right below it, and command cards show the command that ran.

Pull requests and CI fixes use conventional commit messages and exact file paths, with no placeholder logs.

What WorkspaceGPT learns about a repository is stored in files beside your own memory. You can read, edit and remove it, and each entry notes where it came from.

AGENTS.md and CLAUDE.md are read from every workspace root, with a 32 KB budget, and from subfolders when the agent works in them.

Confluence, Jira and Azure DevOps syncs now drop items that were deleted, moved or renamed upstream, so answers stop citing pages that are gone.

When a run or chat gets long, it is summarized and continued instead of running out of room. A startup sync can no longer blank the chat window.

Turn on Debug and the agent gathers evidence first, with log probes and bisecting, before it proposes a fix.

Diagrams render in chat and in spike docs, with zoom and copy. Confluence images can be attached to a page.

Dictate into the composer on macOS with on-device speech (⌘⇧D). Type / for slash commands, and the controls sit beneath a tighter composer.

Bring your Copilot conversations in alongside Claude Code and Cursor.

The desktop runtime dropped an unused model, cutting the installer by about three quarters.

Your choices, such as a Confluence parent folder, are remembered. Spikes check that they covered every source before they finish.

Switch the Sessions panel between Chat and Work. Work groups your sessions by the folder they ran in and opens on a compact home with your tickets and recent activity.

After a pull request is opened, WorkspaceGPT reads its checks and can start a session to fix a failing one. You decide whether it pushes.

Drag the sidebar to the width you like, and zoom the whole app in or out. Settings no longer jump around when you switch pages.

Onboarding offers clearer setup options, and the Knowledge sources you haven't connected show a Connect button instead of hiding.

One dial above the composer sets what the agent may do without asking. Plan can't write at all, and “Run plan” carries a plan out in Agent mode. Manual asks before every change, Auto asks only for risky ones, and Full access, shown in red, asks for nothing.

A terminal panel sits next to the conversation. Commands in code blocks have a Run button, and long-running jobs the agent starts keep going in the background.

Bring your earlier conversations in. Chats from Claude Code worktrees are filed under their own repository.

A session keeps the folder it started in, even if you open a different one later. The Sessions list pages 20 at a time, and a right-click menu opens, pins, renames, groups or deletes a chat.

When the agent finishes it either does the next step, offers it as a question, or lists what's left, so you're never left guessing.

A Connectors page manages Control Chrome. Light and dark themes are remembered across launches, agent tabs open in your own Chrome window, and editing a message keeps its attachments.

On a spike, the agent searches every connected source, writes its findings to a markdown file in docs/spikes, lists its references, and offers a Publish to Confluence button.

Open questions appear as a card you can answer, instead of the agent guessing and carrying on.

Section edits to the same page are batched into a single diff, and pasted edit links are recognised.

Pick where ticket work happens once. A popup confirms when a task would switch folders, and the title bar shows the folder.

GitHub Copilot runs reuse cached prompts, count one request per action, and have an effort picker.

Open Settings → Model and, under “Use a subscription”, pick GitHub Copilot. WorkspaceGPT goes through VS Code's own Language Model API, so it uses the Copilot plan you're signed in to. VS Code asks once to allow it, and requests count toward your Copilot usage. Only models with enough room for an agent run are listed.

GitHub doesn't offer Copilot to desktop apps, so the desktop app signs in the way LiteLLM does: as VS Code's Copilot client. It's off until you click Connect and accept a warning. It can stop working at any time, heavy use can get your Copilot access warned or suspended, and your organization's Copilot policy may not allow it. Sign out from the same card.

Editing and creating Confluence pages, first shipped in Desktop 0.0.6, now works in the VS Code extension too. Reconnect Confluence once to give it edit access.

The agent rewrites one section of a page, or drafts a new page wherever you choose, and shows the change as a diff before anything is saved. Macros, mentions and images outside that section are left exactly as they were, and the page history is your undo. New pages start as drafts unless you ask to publish. Reconnect Confluence once to give it edit access.

Paste a Confluence link and the agent reads the page as markdown, with headings, tables, panels and code blocks intact instead of flattened into plain text.

With the WorkspaceGPT Chrome extension, the agent opens tabs, navigates, clicks, types, fills forms, scrolls and drags. It can also read the page, the console and network requests, and take screenshots. It uses the sites you're already signed in to, acts only in its own “WorkspaceGPT” tab group or the tab you're looking at, and won't type into password fields. macOS for now.

The Sessions list shows only the mode you're in. Chat sessions are listed by date. Work sessions are grouped by project folder, with a + on the folder that's open.

Turn on “Let WorkspaceGPT use this browser” in the Chrome extension, and the desktop agent can list your tabs, read a page and take a screenshot, including pages behind your company's single sign-on. It stays off until you turn it on.

A scheduled sync, or one resumed after a restart, shows its live progress under Settings → Knowledge, and Stop actually stops it.

When a background sync finishes indexing, chat starts searching it immediately. Before this fix, an index built in the background could go unused.

Quit in the middle of a sync and indexing picks up where it stopped the next time you open the app. This now works on Desktop too.

The one-line macOS installer no longer fails with “unbound variable” in a UTF-8 Terminal. Every release is now test-installed from the public download links.

It checks the installer's SHA-256 and installs for your user only, with no admin prompt. It updates itself the same way the Mac app does.

Tokens and keys are stored in Windows Credential Manager, never in a settings file.

Quitting the app ends every command the agent started, even if the app is force-closed.

Open a folder and the agent reads, edits and tests it, with your Knowledge in reach. Runs on Apple Silicon and Intel Macs with macOS 12 or later.

Symbol, definition and reference lookups plus diagnostics, all bundled. It shuts down after five idle minutes to free the memory.

Updates download in the background and are signature-checked. They install when you quit, or right away if you choose Restart Now.

When the window isn't in front, a banner and a Dock badge tell you a run needs your review, has finished, or has failed.

Work grounds answers in your Confluence, Azure DevOps and the open codebase. Chat is a plain conversation that leaves them out. Each mode keeps its own history.

In Remote mode, each answer shows an estimate of the credits it used.