Awesome Jev

Jev is TypeSafe's System One model for probability-based yes/no judgments, choices, and scores rather than generated text.

Check permissions and data handling before running a project. Source review is not a runtime test or a security audit.

Contents

• Getting Started
• Demos
• SDKs and Skills
• Browser and Desktop Tools
• Agent Tools
• Apps and Integrations
• Games and Creative Projects
• Independent Models
• Supporting Drivers

Getting Started

Start with the quick-start example, then build a typed classifier, use Jev in an agent, or explore a local alternative. The source notes record setup, privacy, and evidence limits for the selections below.

Demos

A few examples to watch before digging into the code. These are creator recordings and publisher previews, not independent benchmarks. Linked previews open the original recording or post.

• Jev Ultrafast - A flight-search recording where Jev chooses browser actions and elements; a separate model writes field text. Browser Use · @gregpr07.

• Jev Voice Browser - Partial speech becomes action and target choices, then Playwright operates the page. Moritz Kremb · @moritzkremb.

• jev-align - Label uncertain examples and review proposed classifier or scoring-rubric changes. Sutro. This edits definitions, not model weights.

https://github.com/user-attachments/assets/81650587-e3f1-4655-8213-ed5f6e120e9a

• TypeSafe Computer Use - A Mac desktop recording where OCR and accessibility observations become action choices. @awlevin. Real use controls the mouse and keyboard and needs desktop permissions.

• Jev Browser Skill - A Hacker News walkthrough showing observed controls, bounded choices, and local execution in a small reference implementation. zurfyx.

• tisco - Search video transcripts, inspect matches, and approve clip organization. cairodavila · screenshot published by OpenRouter. This is a screenshot, not a video; Jev judges transcripts rather than video pixels.

The full creator gallery retains the other videos, images, and attribution. Browse the X index, OpenRouter roundup, or computer-use comparison for more examples and limitations.

SDKs and Skills

• TypeSafe Python SDK - Official Python client with a support-ticket routing example.
• TypeSafe JavaScript SDK - Official JavaScript and TypeScript client.
• Official TypeSafe Skill - Request-writing guidance for Noul, Choice, and Score questions.
• Tenbin - MCP server and skill for linting questions, evaluating labeled examples, and choosing thresholds.
• wellposed - Offline request linter for missing options, broken state references, and question types.
• jevrs - Async Rust client with typed answers, derive macros, and native or WASI transports.
• Jevify - Agent skill for finding places to use Jev and planning comparisons with existing code.
• Jev Logs - Scores log entries to decide which need a closer look.
• Skillbox - Self-hosted skill library with optional Jev-based recommendations.

Browser and Desktop Tools

• Jev Ultrafast - Browser-action and element selection with a separate model for text entry.
• Jev Voice Browser - Voice-controlled Playwright browser that chooses actions from partial speech and page controls.
• TypeSafe Computer Use - Mac automation using local OCR and accessibility data, with Jev choosing the next action.
• Cua Driver + jev-use - Python and TypeScript examples where Jev chooses an action, the driver runs it, and code checks the result.
• Jev Social - Bounded social-research loop using Jev choices and socai in the user's Chrome session.
• Jev Browser Skill - Small reference for building and executing browser-action choices.
• Jev Browser - Playwright browser available as an MCP server, CLI, or library.
• Jev Browser Use - Skill using an existing Codex browser connection, with Jev choosing clicks and scrolls.
• Jev Cua - Mac automation through Cua Driver, with optional hosted Jev workflow recommendations.
• jev-bot - Native Mac controls through a JavaScript MCP session, with optional hosted Jev element selection.

Agent Tools

• jev-align - Improves classifier definitions and scoring rubrics using labels and GEPA-proposed edits.
• Winnow - Claude Code context filter that scores tool-output blocks and keeps hidden text locally recallable.
• Jev-Mem - Research memory system where Jev organizes graph memories and guides retrieval while a separate model writes answers.
• patdown - Semantic linter that turns Markdown rules into checks and maps flagged evidence to source lines.
• Oko - Code-search CLI and MCP server with local candidate retrieval and optional hosted Jev ranking.
• jgrep - Semantic matching over code chunks, diff hunks, and CSV rows, published on npm as jevgrep.
• SemDecide - Unix CLI for predicates, choices, scores, JSONL filtering, and CI-friendly exit codes.
• Leanest - Playwright and Vitest test selector with an optional Jev judge and run-tests fallback on uncertainty.
• Supercov - Coverage, security, and code-quality checks where Jev evaluates source files while test coverage stays local.
• Foreman - Checks coding-agent progress and decides when intervention may be needed.
• jev-belay - Claude Code Stop hook that checks local run evidence before judging an unverified completion claim.
• jev-claude-statusline - Claude Code hooks that estimate task progress with Jev and cache it for a status-line display.
• Jev MCP - Tools for checking claims against evidence, screening content, and ranking candidates.
• TypeSafe MCP - Go MCP server for making Jev requests from coding agents.
• jev_search - Experimental file search that ranks candidates by keyword and checks passages with Jev.

Apps and Integrations

• DocJev - Document classification and packet-boundary detection after local parsing or optional cloud OCR.
• tisco - Video-transcript search with previews of proposed clip moves and renames for approval.
• neo4jev - Neo4j graph exploration using edge choices and goal checks inside a bounded beam search.
• MinusPodJev - Podcast ad-detection adapter that scores transcript