WorkspaceGPT Docs

Everything you need to install, configure, and get the most out of WorkspaceGPT — in VS Code, Cursor or Antigravity, or as the WorkspaceGPT Desktop app on macOS or Windows.

Installation

Available in the VS Code and Cursor marketplaces, on Open VSX for Antigravity, and as a desktop app for macOS and Windows.

ext install Riteshkant.workspacegpt-extension

npx workspacegpt

curl -fsSL https://github.com/ritesh-kant/workspaceGPT/releases/download/desktop-latest/install.sh | sh

curl.exe -fsSL https://github.com/ritesh-kant/workspaceGPT/releases/download/desktop-latest/install.ps1 | powershell -NoProfile -ExecutionPolicy Bypass -Command -

Modes & Privacy

WorkspaceGPT has exactly one mode switch, under Settings → Mode. It changes where answers are generated — nothing else.

What each mode sends

Zero data retention

In Remote mode your request is held in memory only for as long as it takes to stream the answer back, then discarded. We do not write prompts, answers, or retrieved snippets to any database, any file, or any log — our servers log status codes and error types only. The entirety of what we store per account is: your GitHub id and handle, your plan and status, an opaque session token that expires in 30 days, and how many credits you have used this week.

Generation itself is performed by an upstream model provider (currently OpenRouter) under its own policy. If you need a guarantee that covers the whole path contractually, use Local mode with Ollama — no third party is involved at all. Full detail in the privacy policy.

AI Providers

These apply to Local mode, where you bring your own model. In Remote mode there is no provider to choose — we run the model for you (see Modes & Privacy).

Configuring Ollama (Recommended)

ollama pull llama3.2:1b

ollama pull llama3.2:4b
# or
ollama pull gemma3:4b
# or
ollama pull mistral

Using your GitHub Copilot plan New

Open Settings → Model. GitHub Copilot sits under Use a subscription, above the API-key providers. Click Connect (or Use once connected), then pick a model. There is no key to paste, and requests count toward your Copilot plan.

Configuring Cloud Providers

Open Settings → Providers, select your provider, and paste your API key. Keys are stored securely in VS Code's secret storage and never logged.

Codebase Exploration

WorkspaceGPT explores the folder currently open in your editor with live search, file reads, and language-server navigation.

Chat, Work & Autonomy

Two switches in the composer decide what a conversation can reach and how much the agent does without asking.

Chat or Work

Pick one before the first message; a conversation keeps the mode it started in. Each mode has its own history, and the Sessions list follows the switch: Chat sessions are listed by date, and Work sessions are grouped by the project folder they were started in, with + on the folder that's open. Sessions from before folders were recorded appear under Other.

How much it does on its own

In Work mode, the permission menu in the composer sets what the agent may do without asking:

Unless you choose Full access, a Confluence edit waits for your approval (see Edit and create pages). When the agent needs a decision from you it asks in a question card instead of guessing, and every run ends with its next step: it does it, offers it as a question, or lists what's left.

Spike tickets

On a research ticket the agent searches every connected Knowledge source, writes its findings to a markdown file in docs/spikes, lists its references, and asks about anything open under Needs your input. A Publish to Confluence button then turns the file into a draft page.

Folders

A chat keeps working in the folder it started in, even if you open another one later. Ticket work can use a default folder you set once, and a popup confirms before a task would switch folders. Git worktrees are supported for branch work.

From ticket to pull request

When a run finishes, Create PR in the git status bar commits only the files the agent changed to a fresh branch, pushes it, and opens your host's new pull request page with the run's report filled in. That works with GitHub, Azure Repos, GitLab and Bitbucket. If the run came from a ticket, the report is also posted on the ticket as a comment. It uses your own git and browser session, so no token passes through the agent.

In WorkspaceGPT Desktop, once the pull request exists the git status bar shows its checks. If one fails, start a session that reads the log and fixes it; you decide whether it pushes.

WorkspaceGPT Desktop

What the standalone app adds around the same agent.

Embeddings & Vector Storage

Connected Confluence, Jira and Azure DevOps sources are turned into vector embeddings so WorkspaceGPT can retrieve the right context. Both halves — making the embeddings and storing them — happen entirely on your machine, in either mode.

Embedding model

Nothing to configure — the models ship with the extension and are used automatically.

Vector storage

Confluence

Connect Confluence with one-click Atlassian sign-in. The agent searches your synced spaces mid-task, reads pages you paste, and can edit or create pages after you approve the change.

Read, edit and create pages New

In WorkspaceGPT Desktop 0.0.6 and later, and in the editor extension from 2.0.45.

• You approve every write. The change arrives as a review card with a before-and-after diff and an Open in Confluence link. Nothing is sent until you approve, and rejecting it with feedback sends the agent back to revise.
• Rich content survives. Macros, mentions, images, status lozenges and complex tables are carried through as placeholders and restored exactly as they were, even inside the section being edited.
• Page history is the undo. An edit saves a new page version, so the previous one is always one click away in Confluence. If someone else edits the page first, the agent re-reads it and redoes its change.
• Never unattended. Runs in Agent mode don't write to Confluence; they put the proposed content in their report instead. Switch to Ask when you want the page updated.

Jira

Connect Jira with the same one-click Atlassian sign-in, so the agent can read the issue it is working on and find related ones.

Azure DevOps

Connect Azure DevOps to chat with work items, user stories, and pull requests.

Browser Control

WorkspaceGPT Desktop can use your own Chrome to check its work: open your app, click through a flow, and read the console and network. It uses your real profile, so pages behind your company's single sign-on just work.

What it can do

Where it stops

• It acts only on tabs in its own WorkspaceGPT tab group, or on the tab you're looking at. Your other tabs are left alone.
• It refuses to type into password fields, including ones inside frames and shadow DOM. Sign in yourself; it then uses the session you created.
• It asks in chat before anything hard to undo, like submitting an order or sending a message.
• While it acts, Chrome shows a “WorkspaceGPT” started debugging this browser bar. Click Cancel there to stop it immediately.

Deployment Automation

Turn the manual release checklist — syncing feature flags, env vars, and component versions across environments — into a reviewable, one-click pipeline. The model is inspired by AWS CodePipeline: a pluggable Source feeds ordered Stages of Actions. Nothing is hardwired to one team's setup.

MCP Server

WorkspaceGPT ships an MCP (Model Context Protocol) server that gives other AI tools — GitHub Copilot, Cursor, Claude Desktop, Claude Code — search over your Confluence, Jira, Azure DevOps and workspace knowledge.

WorkspaceGPT: Connect MCP Server

Commands & Keyboard Shortcuts

All commands are accessible from the Command Palette.

Activity Bar & Title Bar Icons

Reset & Clear Data

WorkspaceGPT stores all data locally in VS Code's global storage. You can wipe everything at any time.

Troubleshooting

Common issues and how to fix them.