cobanov/awesome-jev

Awesome Jev

A curated, source-backed list of projects built with Jev, TypeSafe AI's System One model for fast, typed, probabilistic decisions.

Jev takes program state plus typed questions and returns constrained answers with probabilities. It is designed for software decisions such as classification, routing, scoring, ranking, verification, and guardrails, rather than free-form text generation.

This list favors public source code, concrete Jev usage, clear limitations, and reproducible evidence. The latest review added 52 community projects and resources, bringing that part of the list to 102, alongside official resources, provider integrations, and related lists. See the September 19 research notes for pinned source evidence and review boundaries. Review completed September 19, 2026 (Europe/Istanbul); upstream event dates below are UTC.

Contents

• Start here
• Recent developments
• Official resources
• Provider integrations
• SDKs and developer tools
• Agents, coding, and guardrails
• Context and compaction
• Browser and computer use
• Routing, data, and workflows
• Games, robotics, and interactive demos
• Media and creative tools
• Open reproductions and research
• Evaluation and calibration
• Guides and cookbooks
• Related lists
• Contributing

Start here

• System One shape: text or structured state + typed questions → constrained answers + probabilities → deterministic application code.
• Question primitives: Choice selects an option, Score evaluates ordered rubric levels, and Noul returns a number from 0 to 1 representing the probability of "yes". Review or abstention behavior is defined in application code. See the primitive reference.
• Input boundary: the hosted Jev model is text-only. Browser, audio, image, and robotics projects supply extracted text or structured observations, or use separate perception models. Independent multimodal reproductions are listed separately.
• Good fits: semantic routing, triage, reranking, rubric scoring, moderation, verification, and low-latency decisions inside bounded workflows.
• Important caveat: schema-valid output is not the same as a correct decision. Validate on your own data, calibrate thresholds, keep high-impact actions behind deterministic checks, and provide a human fallback.

Recent developments

• September 18: Python SDK 0.7.0. Release notes document a breaking serialization change from msgspec to Pydantic, a new response_model argument, and corrected serialization of str subclasses.
• September 18: OpenRouter listing. Jev 1.13 is listed with a September 18 date. This is a provider listing date, not evidence of a separate new upstream model revision.
• September 16: Vercel AI Gateway integration. The announcement introduces typed evaluation via AI SDK's experimental evaluate API.
• Current model: TypeSafe documents jev-1.13.0, with both jev-latest and jev-preview currently pointing to it. Pin the version when comparing evaluations.
• Ecosystem refresh: new sections cover context pruning, browser/device control, media tools, and provider integrations. Entries newly added here include both September 18 launches and earlier release-week projects missing from the initial list.

Official resources

• TypeSafe AI - Product overview and early-access entry point.
• Documentation - Concepts, primitives, API, patterns, and SDK guides.
• Introducing System One Models and Jev - Launch post covering the model interface, RLCD, published performance claims, demos, and caveats.
• typesafe-sdk-js - Official TypeScript and JavaScript SDK with inferred answer types.
• typesafe-sdk-python - Official synchronous and asynchronous Python SDK.
• system-one-adapter-python - Drop-in adapter for comparing the System One interface with LLM providers.
• skills - Official agent skills for building and evaluating System One workflows.
• Models and aliases - Version IDs, current limits, pricing, and the distinction between stable and preview aliases.
• Confidence - Distinguishes Choice/Score confidence from answer probability; Noul has no separate confidence field.
• Jev 1.13 jaggedness - First-party limitations covering arithmetic, dates, distractors, adversarial state, and structural inconsistencies.

Provider integrations

• Cloudflare AI - Provider-maintained typesafe/jev integration accepting state and typed questions.
• OpenRouter - Provider listing for typesafe/jev-1.13, alongside the moving typesafe/jev-latest alias.
• Vercel AI Gateway - typesafe-ai/jev through AI SDK's experimental evaluate interface; its Boolean primitive corresponds to TypeSafe's Noul.

SDKs and developer tools

Community-maintained clients and tools; official TypeSafe SDKs are listed above.

• advocaat - Small type-safe client for asking Jev questions about datasets.
• hunch - Probabilistic control flow for Ruby: if Hunch.likely?("fraudulent", given: order) branches on a typed Jev answer, with graded predicates from possibly? to definitely?.
• jev - Elixir/OTP client designed around GenServer replies and pattern matching.
• jev-axi - CLI for picking, rating, checking, ranking, triaging, and guarding from the shell.
• jev-dsl - Early-alpha Haskell DSL that encodes typed question packets and decodes answers; HTTP transport is left to the caller.
• jev-mcp - MCP server exposing classify, score, check, match, and screen tools.
• jev-mcp - An eval-first MCP server for Jev, that returns typed judgments (noul, choice, score) with probabilities instead of generated text.
• jev-shell-history - Ranks existing zsh history entries for inline completion; accepting a suggestion does not execute it.
• jev.nvim - Neovim plugin that splits the buffer into functions with Treesitter, scores each against a plain-language question with Jev, and ranks answers by probability in the quickfix window.
• Jevbridge - ACP/MCP adapter for using Jev alongside coding and chat models.
• jevclient - Async Python client for typed Jev questions and probabilities.
• jevr - Native R client for typed questions and provider-independent answers through TypeSafe or OpenRouter.
• laravel-typesafe-jev - Laravel integration with typed responses, async requests, and testing fakes.
• ruby_decision_model - Ruby client with standard-library transport for TypeSafe and OpenRouter decision endpoints.
• semdecide - Typed semantic decisions for Unix pipelines and CI.
• typesafe-go - Idiomatic Go SDK for the TypeSafe API.
• typesafe-mcp - MCP connector that gives agents access to Jev decisions.
• typesafe-sdk-java - Community Java 17 client for Choice, Score, and Noul, with an optional Spring Boot starter.
• zod-jev - Pairs local Zod shape validation with Jev semantic validation.

Agents, coding, and guardrails

Source-reviewed experiments and integrations. A model judgment does not establish safety or replace the host application's permission checks.

• agent-router - Pre-release Herdr integration that filters eligible coding models by quota and policy before Jev ranks them.
• blink - Navigates file and directory names with Jev-guided walkers to find codebase paths for a natural-language query.
• Canny - Evidence ledger that challenges unsupported "done" claims from coding agents.
• commit-miner - Classifies Git diffs and commit messages into change types and candidate security-fix/CWE labels for inspection.
• foreman - Software-factory supervisor that uses Jev to keep coding agents on task.
• is-malicious - Scans source, configuration, build, and CI files with Jev, then reports suspicious behavior and implicated lines before the code is run.
• jev-agent-skill - Claude Code/ZCode skill that offloads classify, screen, score, and compliance-check judgments to Jev via OpenCode Zen's free tier; ships a retry-hardened zero-dependency caller and a shop comment-triage pipeline.
• jev-belay - Claude Code Stop hook that checks the transcript for evidence before trusting a "done" claim, spending one four-question Jev call only when files changed with no passing check since, and failing open on every error path.
• jev-codex-router - Per-turn Codex model, reasoning, and speed-mode routing.
• jev-commit - Pre-commit hook where one Jev call judges whether the commit message matches the staged diff, flags debug leftovers and unmentioned work, and blocks only when it detects a credential.
• jev-guard - Cross-agent tool-call risk scoring with allow, ask, and deny outcomes.
• jev-pref - Linter that has Jev check code changes against project preferences from jev-pref.json and feeds findings back to coding agents.
• jev-review - Staged code-review workflow with a local dashboard.
• jev-review MCP plugin - Local-first continuous software-quality review for coding agents.
• jev-router - Chooses a model for each fresh Claude Code or Codex turn while wrapping the existing CLI.
• JevRouter - Routes agent requests across models, subagents, skills, MCP tools, CLIs, and plugins with Jev Choice decisions; the host filters by availability, permissions, risk, and confirmation before anything executes.
• jev-scout - MCP server that scores an agent's every search query, result, and fetched page for relevance and credibility, with session budgets, SSRF-guarded fetching, and a live decision dashboard.
• jev-use - Claude Code / Codex / pi plugin where Jev answers batched noul, choice, and score questions and risk-checks tool calls, while a typed escalation contract hands writing and unsure steps back to the LLM.
• jevwire - MCP tools, an embeddable decision library, and advisory or restrictive Claude Code hooks; judgments do not grant native permissions.
• opencode-jev-orchestrator - Keeps an OpenCode parent model fixed and uses Jev difficulty judgments to delegate harder turns to temporary subagents.
• perch - Semantic code linter that evaluates code units against configurable Jev questions.
• pi-jev - Measured tool-call gate and general typed decision layer for the Pi coding agent.
• pi-warden - Pi extension that judges rule compliance, risky actions, stuck loops, and completion claims; enforcement depends on the hook and policy.
• skillbox - Self-hosted skill library with optional Jev relevance recommendations over an authorized catalog.
• skillranker - Rust CLI that ranks agent skills against live session context and can abstain.
• supercov - Scores source files so coding agents can prioritize code-quality work.
• wakegate - Experimental gate where Jev decides whether a timer or incoming event is worth resuming a sleeping agent's LLM; code skips only when Jev is confident and always wakes on user messages, errors, and a skip limit.

Context and compaction

These tools select what reaches a model. Preserving retained text verbatim does not prove that omitted history was unnecessary.

• fast-dev-compaction - Codex port that restores Jev-selected verbatim history around native session compaction.
• fast-jev-compaction - Claude Code plugin and library that score tool-call/result pairs for deletion or truncation while retaining selected text verbatim.
• pi-fast-jev-compaction - Pi extension that prunes stale tool history and leaves summary compaction to Pi when pruning is insufficient.
• pi-jev-compact - Selective, verbatim context compaction for Pi using Jev model.
• pi-jev-context - Opt-in Pi extension that filters older messages from model requests while preserving the original session history.
• winnow - Judges tool results before admitting them into Claude Code context.
• yoshi - Experimental Claude Code/Codex proxy that uses Jev to prune request context while preserving tool-call protocol structure.

Browser and computer use

These projects can operate real browsers or devices when enabled. Published demos have task-specific success criteria and do not establish general reliability.

• BrowserClaw - Zero-lock, session-preserving Chrome MCP server that couples a local Jev System One semantic micro-loop (chrome_act_toward_goal) with an 85%+ pruned DOM tree (Shadow DOM & iframe pierced), dispatching native CDP events (isTrusted: true) on active logged-in sessions without focus theft.
• jev-browser - Browser-use experiment powered by Jev decisions.
• jev-browser-use - Codex browser skill that uses Jev for navigation and target selection while Codex handles text entry and outcome verification.
• jev-social - Local Instagram and TikTok research app where Jev selects a platform and bounded socai operation from observed state, while code enforces confidence and the socai CLI executes the browser step.
• jev-ultrafast - Browser agent where Jev selects an operation and compatible DOM target, and a separate LLM supplies typed text.
• jev-voice-browser - Maps partial speech transcripts to browser intents and observed targets, with code deciding whether to act, wait, or ask.
• Jev for Chrome - Unofficial Chrome extension (Manifest V3) port of Jev Ultrafast: Jev picks the operation and DOM element in one request, a small text model writes typed values, and it runs in the user's own tabs through OpenRouter, TypeSafe or Cloudflare; includes a 17-task headless-Chromium suite with recorded traces.
• mobile-jev - Android agent using Mobilerun observations and bounded Jev actions; execute mode controls a real device, while the published Uber demo stops before booking.
• typesafe-computer-use - macOS computer-use experiment using OCR plus bounded Jev action selection.

Routing, data, and workflows

• duckdb-jev - DuckDB extension that exposes Jev judgments as SQL values with return types derived from the declared criteria.
• HA-Jev - Home Assistant integration exposing typed answers as sensors, automation actions, and an Assist conversation agent.
• hono-jev-router - Routes Hono HTTP requests by meaning.
• jev-align (Sutro) - Active-learning CLI where Jev evaluates rows and surfaces uncertain or audit samples for human labeling, then GEPA proposes revised decision definitions that the user can accept or reject.
• jev-curate - Streaming filter and scorer for Parquet and JSONL datasets.
• jev-logtriage - Jev scores collapsed log batches for noise, severity, and whether an operator should act. Code maps the answers to suppress, watch, review, notify, or page. Nothing is executed.
• jev-reranker - Retrieval and RAG: uses Jev Noul judgments to assess retrieved documents for relevance and usefulness as answer evidence, then sorts results and optionally filters them using a configurable threshold.
• jev-reviewer - Research-document extraction aid where Jev selects and verifies source lines for verbatim quotes; findings require human review and are not clinical decisions.
• jev-search - Uses Jev to select search sources and rank Search1API results, returning source links and snippets.
• jev-trade - Hyperliquid trading desk where Jev answers Choice questions for long/short, open/close/hold, and leverage; application code quotes or sends no order. Defaults to a dry run; a live key can place real orders.
• jev-trader - Kuru/Monad trading experiment with optional Jev buy/sell decisions; defaults to a mock model and dry-runs without a private key, but configured execution can place real orders.
• jev-tree - Recursive choice over taxonomies larger than Jev's direct option limit.
• Jev Web Analyzer - Fetches a public SaaS page as Markdown, then asks Jev ten bounded questions about likely first-visit comprehension; URL validation, caching, sanitization, and presentation remain deterministic. Demo.
• jevlogs - OpenTelemetry log triage before more expensive analysis.
• jevql - psql-shaped client and Go/TypeScript/Python SDKs for vanilla Postgres where Jev makes Noul, Choice, and Score judgements about individual table rows after the plain SQL has run on the server, and the client applies the resulting filter, sort, or group.
• jevsql - SQL-like filtering, ranking, classification, and scoring with natural-language predicates.
• llama-index-jev - LlamaIndex reranker and selector using Jev Score and Choice answers, with configurable confidence handling.
• n8n-nodes-typesafe-jev - Community n8n node for asking multiple typed questions over workflow state.
• pg-jev - PostgreSQL extension for semantic questions over table rows.
• pg_typesafe - Pre-alpha PostgreSQL C extension exposing Choice, Noul, Score, and batched judgments from SQL.
• tax-doc-classifier - Classifies text-bearing PDF pages into IRS form and page-kind candidates with a confidence gate; document triage, not tax advice, and scanned pages need OCR.
• tiershift - Policy-bounded model routing for TypeScript and Python.
• typesafe-jev-workflow - LangGraph email-intent workflow using a typed Jev choice.

Games, robotics, and interactive demos

• heist-one - Browser stealth game where Jev judges guards while deterministic code owns the world.
• jev-canvas - Voice and finger-pointing control of a tldraw canvas: Jev picks the action, target shape and place from each partial transcript plus the fingertip position; deterministic code applies thresholds and executes the edit.
• jev-drone - Simulated MuJoCo quadrotor with Jev making slower tactical judgments from processed camera observations; deterministic code controls flight.
• jev-experiments - Collection of inspectable Jev demos, including scripted support conversations with typed intent, escalation, and suggested-response decisions.
• jev-palette - Command-palette demo where a single Choice question over a 77-command catalog turns the returned probability distribution into the per-keystroke ranking for Portuguese or English input; deterministic code executes the selected command, with a classic fuzzy-match baseline shown side by side.
• jev-plays-pokemon-red - Pokemon Red on PyBoy where deterministic code owns the route and arithmetic and Jev picks only at branches, with every battle turn's faint prediction scored by Brier against the emulator's RAM state.
• jev-tetris - Tetris where deterministic code enumerates every reachable placement, including tucks and spins, and writes each one as an English sentence; Jev returns a probability for all of them across five Choice questions, and the highest-weighted option is played while the whole distribution is drawn on the board. Ships a shuffled-probability control and a 23-line regex baseline over the same sentences, which outscores the model.
• [JevPilot](https://github.com/standardagents/je