Desktop Agents: hand tasks to Claude Code, Codex and more

Agents in the WidgetStore desktop app is a control plane. It doesn't bring its own AI: it hands your task to the AI coding agents already installed on your computer, runs them in their official headless mode, and turns the output into a readable timeline.

Supported agents

Agent	Command	Headless mode we use
Claude Code (Anthropic)	claude	claude -p --output-format stream-json
Codex CLI (OpenAI)	codex	codex exec --json
Gemini CLI (Google)	gemini	gemini -p --output-format stream-json
Antigravity CLI (Google)	agy	agy -p --output-format stream-json
Cursor Agent	cursor-agent / agent	agent -p --output-format stream-json
GitHub Copilot CLI	copilot	copilot -p
opencode	opencode	opencode run --format json
Aider	aider	aider --message

The Claude, ChatGPT, Codex and Cursor desktop apps have no headless mode: Open copies your prompt first so you can paste it in the app.

Open Agents → Agents in the sidebar to see what was found, the versions, and whether you seem to be signed in (we only check whether a sign-in file or environment variable exists — we never read the secret itself). Missing ones link to install instructions; choose Rescan after installing.

Run a task

1. Choose New task (or type “Ask Claude Code: …” in the launcher).
2. Pick the agent and the folder it works in (recent folders are listed).
3. Describe the task, choose permissions, then Run (Ctrl + Enter / ⌘↵).

The task shows up under Tasks: what the model says, the tools it calls (commands, files read or written) and their results, diffs, elapsed time and tokens, all live. You can Stop at any time. Choose how many run at once (1–4); the rest wait in a queue. You get a notification when a task ends, and the tray tooltip shows how many are running.

Permission profiles

Profile	Meaning	Agent flags (examples)
Read-only (default)	Looks and plans only; no file changes, no commands that change things	Claude --permission-mode plan, Codex --sandbox read-only, Cursor --mode ask
Edit files	May edit files in the folder; other commands are still denied by the agent’s own rules	Claude acceptEdits, Codex workspace-write, Gemini auto_edit
Auto-approve everything	Never asks: any file, any command, network	Claude --dangerously-skip-permissions, Codex --dangerously-bypass-approvals-and-sandbox

• Each agent only offers the profiles it supports (opencode has no auto-approve, for example). We never silently fall back to a wider profile.
• Auto-approve needs a separate risk confirmation on every run.

Safety and privacy

• Nothing runs until you choose Run on your computer. The website (for example “Run with agent” on an SOP step) can only prefill the New task window, which says where it came from. Website prefills can’t pick auto-approve or a folder.
• Requests from the command line or other agents through the WidgetStore CLI open a confirmation window showing the agent, folder, prompt and permissions; it runs only if you choose Allow.
• Agents get only the environment variables the system needs plus that agent’s own settings. Our sign-in and tokens are never passed to an agent.
• Prompts, output and results stay on this computer and are never uploaded. Delete one task or clear them all under Tasks.

Example templates

The New task window and the Agents page include a few examples: a Sentry sweep (needs the Sentry MCP server set up in your agent), reviewing open PRs, release notes from git log, a 10-line tour of a codebase, and running tests and fixing failures (this one defaults to Edit files). Click one to prefill, adjust, then run.

Usage

When the CLI reports tokens (Claude Code, Codex, Gemini, opencode, Aider), each finished task shows them; Claude Code, opencode and Aider also report their own cost estimate (subscriptions aren’t billed per token, so treat it as a rough guide). For long-term usage across tools, open AI usage in the settings.