Agent access: let Claude Code, Codex and Cursor use the desktop app
Let Claude Code, Codex, Cursor or Claude Desktop use the WidgetStore desktop app over local MCP and the CLI: one-click connect, permissions, confirmations for risky actions and an undoable log.
Everything the WidgetStore desktop app can do on your computer can be used by your own AI: Claude Code, Codex, Cursor, Claude Desktop and others call it through local MCP or the widgetstore CLI. All capabilities come from one list, so new ones show up for your AI automatically.
Connect in one click
Open the desktop app → Settings → Agent access:
- Under “Connect in one click”, click Connect next to Claude Code / Codex / Cursor / Claude Desktop.
- The dialog shows exactly which file changes and what is added, and lets you pick a permission: Read only, Read & write, or Include risky actions (you confirm each).
- Restart that AI tool and it’s ready.
Each connection gets its own token, written only into that tool’s config file. Nothing else in the file changes, and a backup (<file>.widgetstore-backup) is kept before the first edit. Under “Connected clients” you can change the permission, make it ask again, or Revoke it (the token stops working at once and the entry is removed from the config).
What your AI can do
- Read: desktop status, system monitor, launcher search, focus status, today’s to-dos, the Agents board / background tasks / finished tasks, Loops, desktop widgets, the pet, now playing, unread message count, Reader feeds and articles, model gateway models and usage, AI usage.
- Private read (each client asks you once in the desktop app; you can remember it): My day timeline, screen time, recent conversations, email lists and bodies.
- Write: start / stop focus, add / check off to-dos, show notifications, open pages, pin widgets, turn Loops on or off, control music, subscribe to feeds, mark emails and articles, start a note, pause screen time.
- Risky (you confirm each time in the desktop app; you can remember it per client): run an agent task, message a friend, send an email, unsubscribe a feed, run a terminal command.
- Terminal commands: the confirmation shows the full command and folder; you can remember a command prefix (remembering
git statusdoesn’t covergit push). Commands run in a clean environment without the app’s sign-in or keys, time out after 60 seconds by default, can be stopped, and output is cut off. They really run, so they can’t be undone; Activity records the command, folder and exit code, not the output. - Never offered: passwords, the vault, one-time codes, cookies, API keys or tokens.
The full list with arguments: run widgetstore help --json or widgetstore docs. The app also writes AGENTS.md and llms.txt to its config folder (Settings → Show in folder).
Manual setup
- stdio:
widgetstore mcpwith theWIDGETSTORE_AGENT_TOKENenvironment variable (Connect sets it for you). Programs that aren’t connected can’t call anything and are told to connect in Settings → Agent access. Connecting Claude Code uses its ownclaude mcp add-json --scope user. - HTTP: the app serves MCP on a random port on
127.0.0.1; the URL is inagent-mcp.jsonin its config folder (only you can read it). SendAuthorization: Bearer <the token from Connect>; web pages in a browser can’t reach it. - CLI:
widgetstore <capability> --json. Exit codes: 0 ok · 1 error · 2 usage · 3 not allowed or turned off · 4 desktop app not running · 5 you declined. See WidgetStore CLI.
Activity and undo
Every call (succeeded, failed or declined) is listed in Settings → Agent access → Activity, kept on this computer only, with password-like fields removed. Actions that can be reversed (checking a to-do, turning a Loop on or off, subscribing, marking emails, an email still in the undo-send window, removing a desktop widget, pausing screen time, starting focus) have an Undo button.
The main switch at the top stops every local call at once.
Last updated: Nov 18, 2026