What does this skill do, and when should you use it?
jev-use is a recipe defined in skills/jev-use/SKILL.md of the trycua/cua repository (MIT licensed). Its core design keeps the decision layer entirely above Cua Driver: the Driver supplies observations and executes actions, the application constructs complete candidate tables, and TypeSafe Jev returns only one candidate ID — it never invents tool names, coordinates, or arguments. The recipe prescribes strict rules for observation freshness, visual evidence validation, native accessibility candidate construction, and credential handling, and offers a deterministic mock path so development can proceed without a model API key. A runnable reference lives in libs/cua-driver/examples/jev-use/.
- Obtains fresh Cua Driver observations (browser DOM, native accessibility tree, or visual regions) through one persistent CLI or MCP session.
- Builds a candidate table capped at 24 action candidates plus reobserve and abstain, each with the complete Driver tool and arguments.
- Communicates with TypeSafe Jev over stdin/stdout using the cua.jev_choice_request_v1/v2 protocol, sending only goal, compact observation, history, and candidate IDs.
- Validates the returned choice, rejecting unknown, duplicate, stale, capture-mismatched, or below-confidence results.
- Executes at most one Driver action (background delivery by default), then reobserves and verifies completion against an independent postcondition.
- An engineer building a desktop automation agent who wants the model restricted to a whitelist of application-owned actions instead of freely generating coordinates and tool calls.
- A team that wants System 1 style selection decisions over structured interface elements, with execution firmly locked behind driver boundaries.
- Developers driving native desktop apps on macOS, Windows, or Linux who need candidates built from the accessibility tree (AX/UIA/AT-SPI).
- Automation scenarios mixing browser DOM with optional visual grounding, where every visual click must be bound to an exact capture ID.
- Teams that want to develop against the deterministic mock path first, then switch to live Jev without code changes in the recipe.
- Use cases where the model should freely generate coordinates, refs, or tool calls — the recipe explicitly forbids Jev from inventing any arguments.
- Users without a running Cua Driver environment (macOS/Windows/Linux desktop) or unwilling to integrate a TypeSafe Jev adapter.
- Tasks that require actions beyond the default risk policy — delete, send, purchase, and close actions are excluded unless the task spec allows them.
How do you install this skill?
- This is a static source review; no tests were executed. Key-path reproducibility is inferred from file self-consistency and CI configuration.
- The skill depends on external Cua Driver and extension contracts (including AGPL-licensed cua-perception assets); verify licensing and trust boundaries before deployment or redistribution.
- Reachability of core dependencies (cua.ai installers, GitHub, Hugging Face) from mainland China is not declared; Chinese users may need extra network configuration.
- Foreground delivery and unrestricted permission mode (CUA_DRIVER_PERMISSION_MODE: unrestricted in CI) carry real desktop-action risk; production use should enable approval and authorization flows.
- Publisher identity is unverified/unknown; the skill file itself lacks a version number and update path.
- Shell / CLI
- Network access
- Local filesystem
- MCP Server
Cua DriverTypeSafe Jev adapter (mock path runs without TYPESAFE_API_KEY)
The source documents no per-host install command for this skill alone. The repository README notes that the cua installer can install cua skills into AI coding agents:
curl -fsSL https://cua.ai/install.sh | sh
cua auth loginThe installer offers to add cua skills and the cua MCP server to Claude Code, Codex, Cursor, and others; the skill file itself is at skills/jev-use/SKILL.md in the repo.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Set up a bounded computer-use loop in mock mode following the jev-use recipe that clicks a button in Calculator and verifies the result.
- Build a candidate table for my desktop automation agent so the model can only choose from application-owned candidate IDs, never generate coordinates.
- Following libs/cua-driver/examples/jev-use/, wire up NativeAccessibilitySource and cua.jev_choice_request_v2 for a native desktop app.
- How should jev-use rules trigger reobserve or abstain when visual and accessibility evidence disagree?
The skill is triggered as a recipe/guide describing how to structure the jev-use loop: state the goal and get a fresh Driver observation through one persistent CLI or MCP session; prefer fresh accessibility or browser DOM tokens; if visual grounding is needed, discover parse_visual_regions through the current MCP tool inventory and validate its versioned result and capture ID; build a bounded candidate table and send only goal, compact observation, history, and candidate IDs to Jev; resolve one returned ID, execute at most one action, and reobserve. Browser tasks use the v1 protocol; native desktop apps use NativeAccessibilitySource and v2. The runnable example is in libs/cua-driver/examples/jev-use/; use checked-in fixtures for deterministic development. The mock path requires no TYPESAFE_API_KEY; the live key comes only from the process environment or a secure interactive prompt.
What are this skill's strengths and limitations?
- Crisp decision boundary: the model can only pick from application-constructed candidate IDs, eliminating hallucinated arguments and unbound coordinate clicks.
- Deterministic mock path works with no model API key; Python and TypeScript adapters expose equivalent mock and live behavior.
- Explicit, actionable rules for freshness, visual evidence validation, native role mapping, and credential handling.
- Ships a runnable example and checked-in fixtures; the host repository is MIT licensed.
- Tightly coupled to the Cua Driver and TypeSafe Jev ecosystem; unusable outside that stack.
- Candidate table capped at 24 actions and delete/send/purchase/close excluded by default, limiting flexibility.
- The visual adapter only works when the Driver advertises both parse_visual_regions and the capture-bound click.capture_id input.
- The source provides no independent test suite or benchmark results proving the recipe's own success rate.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| jev-use — A Bounded Computer-Use Loop over Cua Driver this page | 58 · Recommended | ★ 29k | 1d ago | MIT |
| Desktop App Control with Computer Use | 61 · Recommended | ★ 6.4k | 3d ago | Apache-2.0 |
| squirrelscan Website Audit & Fix Loop | 49 · Use with care | ★ 97 | 3d ago | MIT |
| Angular Developer Skill | 63 · Recommended | ★ 674 | 3d ago | — |
| SwiftUI WCAG Accessibility Auditor Skill | 58 · Recommended | ★ 13 | 7mo ago | — |
The repo README covers the whole Cua product family (Driver, Spaces, Lume, CUA-S1, Bench); jev-use is one specific recipe for integrating a TypeSafe Jev decision model with Cua Driver, as opposed to letting a general agent freely call Driver tools. The source names no third-party alternatives.
How did FollowSkills review this skill?
SKILL.md enforces least privilege and clear safety boundaries: Jev may only select from application-owned candidate IDs, never invent tool names or coordinates; capture_id binding prevents stale coordinate clicks; delete/send/purchase/close actions are excluded by default; keys are read only from environment or secure interactive prompts, never in source, args, or logs; foreground delivery requires explicit escalation/authorization. Deductions: trust relies on external Driver/extension contracts (e.g., cua-perception from signed release assets with AGPL caveats) and the file itself lacks detailed rollback and user-confirmation flow.
The decision loop, freshness rules, and failure modes (reobserve/abstain, rejecting unknown/duplicate/stale choices) are clearly specified and self-consistent; the mock path works without an API key and a dedicated CI workflow (authorized-live-jev-use.yml) exercises mock checks. Deduction: static review only — the referenced example directory libs/cua-driver/examples/jev-use/ was not provided, so key-path reproducibility and failure-feedback quality cannot be confirmed; capped at 10.
Trigger conditions are precise in the description (jev-use recipe or similar Jev integrations; explicitly not for adding model logic or credentials), with declared boundaries between browser v1 and native v2 and stated non-fit ranges. Deductions: narrow developer audience requiring TypeSafe Jev and Cua Driver; no Chinese-language documentation; reachability of core dependencies (cua.ai installers, Hugging Face, GitHub) from mainland China is not addressed.
Well-layered structure (decision loop, freshness, native candidates, credentials/proof), MIT license is clear, and the repo shows version-anchoring and governance signals. Deductions: the skill file itself has no version, changelog, or maintenance-owner statement; install/dependency notes live outside the skill.
A runnable reference example and deterministic mock path are provided, with verification via an independent application oracle rather than model output; the decoupling-of-decision-and-execution value proposition is clear. Deductions: static review cannot confirm outputs are directly usable; example/test details are only partially visible and comparative-benefit evidence is limited.
Auditable primary material exists: specific contracts in SKILL.md are cross-corroborated by repository CI workflows including mock assertions, credential-leak checks, and audit scripts. Deduction: nothing was executed, coverage is limited, and independently reproducible conclusions are not achievable from a static read.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →