Is this usable without a ChatGPT subscription or OpenAI key?

Yes. Jev-only guided mode needs only a TypeSafe key and is the default. Hybrid mode adds subgoals and text drafts from the GPT-6 Astra planner and therefore needs an OpenAI Platform API key, which a ChatGPT subscription does not include. Fill a form needs no provider connection at all.

Can Otto click something it decides on its own?

For a guided task, no. Otto presents the exact action before executing it and the next observation checks what changed, and you confirm the final result. The displayed confidence is labelled a model score rather than a guarantee. The planner cannot invent unrestricted coordinates or shell commands.

Should I install a downloaded build on my work machine?

Read the release notes first. Current macOS packaging is ad-hoc signed without Developer ID signing or notarization and Windows publisher signing is not configured, so neither platform gives you a trusted publisher. For source builds you need Node.js 24+, npm, and macOS 14+ with Xcode command-line tools or Windows 10/11 with PowerShell 5.1.

What can the agent bridge actually reach?

A source build exposes scoped native inspection, actions and verified workflows, and run_steps makes no model calls while checking supplied values against native state. It covers native press, fill and vertical-scroll controls plus Enter, Escape and Tab. It does not yet provide arbitrary visual grounding, drag, app launch or a browser DOM driver, and it is read-only unless started with --allow-actions.

How do I run Otto from source?

Node.js 24+ and npm, plus macOS 14 with Xcode command-line tools or Windows 10/11 with PowerShell 5.1. Clone the repo, npm install and npm run dev; if your package manager disables dependency scripts, run node node_modules/electron/install.js.

Which Otto workflow runs without any API key?

Fill a form. It takes a literal field map for one app, reviews all before and proposed values once, checks each write and the final field set through native readback, issues no submit, and makes no model calls.

Where do Otto's API keys live, and is a saved key verified?

Keys stay in the main process and optional persistence uses Electron's OS-backed encryption; session-only keys are re-entered after a restart and the renderer never receives stored keys. A saved TypeSafe key is not verified until the first guided task makes the live request.

CommunityJev Demos & Experiments

Otto

Open-source Electron computer-use agent for macOS and Windows that reads the apps you select through native accessibility with local OCR added for text targets, and asks Jev to choose concrete actions that it shows you before executing.

See highlights

Open github.com

Stars: 4
Language: TypeScript
License: MIT
Published by: Community
Author: NobleSpartan6
Last verified: Oct 2026

NobleSpartan6/otto

User repository on GitHub

View on GitHub

Open-source native computer use for macOS and Windows: TypeSafe Jev, local OCR, and selective planning.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

An open-source desktop computer-use agent for macOS and Windows. Otto combines TypeSafe Jev decisions with an optional GPT-6 Astra planner, and can apply an exact set of form values after one review without calling a model. It works through native accessibility controls and local OCR.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Four workflows: guided Jev-only, guided hybrid with a GPT-6 Astra planner, Fill a form with no model calls at all, and a source-build MCP agent bridge.
Only explicitly selected running apps are observed, and Otto's own UI plus protected system dialogs are excluded.
Fill a form takes a literal field map, reviews all values once, checks every write and the final field set through native readback, and issues no submit.
OCR runs locally with bundled English language data and never downloads models at runtime; password controls are excluded from actions and observations.

Watch out

This is an early development alpha. Broad app compatibility, model-driven task success, trusted publisher signing and subscription sign-in are all unverified or unfinished. macOS packaging uses ad-hoc integrity signing without Developer ID signing or notarization, and Windows publisher signing is not configured, so Gatekeeper prompts remain. On Windows access is limited to the normal user desktop; elevated apps and UAC screens are unsupported. Hybrid mode needs an OpenAI Platform API key, and a ChatGPT subscription does not provide API credits. Canvas and icon-only interfaces remain a limitation. A saved TypeSafe key is not verified until the first guided task makes the live request. Stop cancels future dispatch but cannot undo input already delivered. Guided-task exports omit screenshots and raw observations, yet task descriptions and fill receipts can still contain user-entered data. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

ui-generator-instinct-jev

Demo that turns free-text UI descriptions into real shadcn components using only Jev typed Choice, Noul, and Score answers, with no generated code on the decision path.