It works across files, the terminal, browser and apps on your Mac—planning, acting and checking the result. You decide when it must ask first.
Code changes, documents, browser actions and scheduled jobs move through one agent. Follow its steps and inspect the work it produces.
Read, edit, run, verify.
Point genduo at a folder. It reads the code, edits files, runs your tests in the shell and keeps going until the job is done. Sensitive commands and writes outside the workspace stop for your approval.
See how approvals workReports, tables, pages, slides.
Ask for a weekly metrics digest and get a real document with tables and charts instead of a wall of chat. When it is ready, publish it to a genduo.app link you can hand to anyone.
What you can deliverOpen, click, type, screenshot, verify.
On the Mac app, genduo can open pages and fill forms in a real Chrome window, take screenshots, and click, type and launch apps on your desktop. Screenshots and reads run freely; actions that change something ask you first.
Get the Mac appCron, in your time zone.
Turn any task into an automation with a cron schedule and a time zone. Runs keep the same boundaries as any other task: sensitive actions are refused rather than guessed, unless you explicitly authorize the automation to run unattended with full access.
Read the FAQSet your standards with skills, connect tools through MCP, carry context forward with memory and delegate bounded jobs to sub-agents.
A skill is a SKILL.md file: a little frontmatter plus the instructions you want followed. Import one from a URL or write your own, then call it with /name inside any task.
Connect any MCP server over HTTP or stdio. Its tools appear as mcp__server__tool, and large catalogs load lazily through mcp_list, mcp_get and mcp_call so your context stays small.
Hand a bounded job to a sub-agent with its own context window. It shares the parent's approval boundary, so nothing gets past you through delegation, and it never runs a pricier model than the one you picked.
genduo keeps titled notes it can search later: your conventions, decisions and the quirks of your setup. Anything that looks like a credential is refused, so memory never becomes a secret store.
Pick an approval level per task. genduo asks before the actions that matter, shows you exactly what it is about to run, and refuses what you have not allowed.
Questions about safetyUse built-in models, let Auto choose, or connect your own OpenAI-compatible endpoint. The same tools and approval controls work with each option.
Leave the selector on Auto and genduo scores each request: long messages, large contexts and recent tool chains route to a heavier model, quick questions stay on a fast one.
Set how much a model should reason for a task or a whole session, turn it off for quick edits, or keep the provider default.
Add any OpenAI-compatible endpoint with your own key. On the Mac app the key lives in ~/.genduo on your machine and is never sent to genduo servers.
A delegated sub-agent runs on the same model as its parent or a cheaper one. It can never quietly escalate to a pricier model.
The web app runs the same agent without shell or file access: research, writing, documents and sharing. When you want it inside your codebase and your apps, switch to the Mac app.
Sign-in, model access, billing and published links are shared between the web app and the Mac app. Tasks you run on your Mac, and their files, stay on your Mac.
In the browser genduo reads the web, writes documents and generates images, but it has no shell and no access to your files. Those tools exist only on your Mac.
Ask for a table, a page, a deck or an image and get a real artifact, not a description of one. Publish anything to a genduo.app link when it is ready.
Tables with real cells you can sort and edit, built from research, files or exports.
Landing pages, product pages and small multi-page sites, published on a link.
Decks with a consistent layout from an outline, a document or a conversation.
Illustrations and cover images generated to fit the document they belong to.
Short answers to the questions people ask before their first task.
A Mac with Apple Silicon or Intel for the desktop app, or any modern browser for the web app. Sign in, point genduo at a folder if you want it working in your files, and start a task.
Claude, GPT, Gemini, DeepSeek, Qwen, GLM and MiniMax models are built in, and you can add any OpenAI-compatible endpoint with your own key. Leave the selector on Auto when you would rather not choose.
Keys you add yourself are stored on your Mac under ~/.genduo and are never sent to genduo servers. Tasks you run on the Mac app, and the files they touch, stay on your machine. The cloud handles sign-in, model access, billing and the links you choose to publish.
You choose per task. The default asks before sensitive commands and before writes outside the workspace; a stricter level asks before every write or network operation; full access skips ordinary prompts. At every level, .git directories and .env files are never written, and full access still refuses writes outside the workspace.
Yes. Add any OpenAI-compatible endpoint, including one running on your own machine, and pick it per request. Sign-in still needs a connection.
genduo is not an editor plugin. It is an agent that works across your whole Mac: repositories, documents, the browser and desktop apps, on a schedule if you want. You review and approve instead of typing alongside it.
Start in the browser, or give genduo access to your files, terminal and apps with the native Mac app.