Jev directory
CommunityJev Integrations & Adapters
Jev in Yappy
Yappy's macOS computer-use agent runs every step through Jev: the click, field and scroll are chosen from typed decisions over an accessibility text table, with no screenshots sent and a chat model called only when a step needs words.
- Reading time
- 3 min
- Published by
- Community
- Author
- Yappy
- Last verified
- Oct 2026
- Platforms
- macOS
Highlights
- Available in Yappy 0.4.60 and later; it needs macOS 13 or later on Apple Silicon, and the lane is on by default on the Yappy plan.
- Jev reads the front window through macOS accessibility as an indexed table of controls — role, label, value, position — so no screenshots are sent.
- One operation question (click, type, press Enter, scroll, wait, done, blocked) plus one target question per operation, answered before anything moves.
- In the measured run, five Jev decisions took 275–690 ms each and cost $0.003; the whole job application took 1m 54s against 13m 14s for the competitor.
- Hands off to the full agent when no window is readable, confidence falls under 0.25, three actions change nothing, or a needed value is missing.
Watch out
Grounded in the yappy.biz/jev page read on 2026-10-15, including Yappy's own benchmark from 18 September 2026 — one run per column, with method and caveats on its linked research page. Yappy is a commercial macOS agent and Jev calls are billed to the Yappy plan with no separate key.
Questions about this entry
Short answers drawn from the linked project or page.
Yappy is a computer-use agent for macOS: you say what you want and it does it in your existing apps. On the Yappy plan every step now runs through Jev, TypeSafe's System One decision model, with the full model called only when a step needs words.
No. Yappy reads the front window through macOS accessibility as a text table of controls — role, label, value and position — and that table is the state Jev decides over. No image of your screen is sent to Jev.
In Yappy's measured run on 18 September 2026, each Jev decision took 275 to 690 milliseconds, and five decisions together cost $0.003. TypeSafe quotes 70 to 500 ms end to end and $0.042 per million input tokens with output free.
Same job application, same answers, same do-not-submit instruction: 1m 54s end to end versus 13m 14s, 22 actions versus about 104, 0.39M input tokens versus 11.24M, and about $0.24 versus $4.24. One run per column, per the page.
Set "fastLane": false in ~/.yappy/config.json. The lane is on by default for every session on the Yappy plan, and turning it off is the documented opt-out.
It hands off to the full agent in the same session with a note of what was already done. The documented triggers are no readable front window, confidence under 0.25, three actions in a row that changed nothing, or a field needing a value the task did not give.
Jev writes no prose and cannot run an agent turn on its own; a small chat model supplies text when a step is type. Jev calls are billed to the Yappy plan like the rest of the agent's inference, with no separate key or configuration.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Voice-controlled computer use on a Mac
X: Computer use without screenshots
X: Copy and paste, with a decision in between
X: Jev driving real Chrome from Grok Bot
More Jev resources like this
A local Codex and Waku plugin for app-agnostic Mac computer use: Accessibility actions with a visual fallback, approval-gated mutations, and a Jev fast path for repetitive tasks.
Jev Agents & MCP#community#catalog-import#madewithjev
Community
eronmmer/jev-cuaTypeScript
eronmmer/jev-cuaTypeScriptmacOS menu bar screen-time app that reads the screen as text through the Accessibility API and asks a System One model one Choice per user rule plus a Noul on whether the page is a payment, login or banking screen, stepping in only when a rule's probability clears its threshold.
Jev Integrations & Adapters#community#macos#digital-wellbeing
Community
RoderickQiu/qualmPython
RoderickQiu/qualmPythonAn open-source computer-use MCP server for Claude Code, Codex and other agents that reads the accessibility tree instead of screenshotting every step, using Jev as a fast judge and, experimentally, as image input to spot icons and game objects.
Jev Agents & MCP#community#github#mcp
Community
8itlab/reflex-cuPython
8itlab/reflex-cuPythonA governed execution layer and drop-in proxy between typed-decision models like Jev and callers such as Claude Code, Cursor/MCP and SDKs: raw probabilities become deterministic policy actions from ALLOW to DENY with a DuckDB audit trail.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
0xshikhar/jev-fusePython
0xshikhar/jev-fusePythonA fork of codex-with-chatgpt that adds TypeSafe Jev plus deterministic rules to decide when Codex hands planning, debugging or review to web ChatGPT and when it switches back. Smart switching is off by default, and with it off the tool behaves exactly like the original.
Jev Integrations & Adapters#community#catalog-import#logicrw
Effortlane is a local router for Codex on macOS that picks model and reasoning effort with Jev before each user turn, keeps native ChatGPT authentication, and records routing evidence locally. Shadow mode logs recommendations without changing execution.
Jev Integrations & Adapters#community#catalog-import#logicrw
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.