Launches
Products built on Jev, launched by their makers once a week. People vote with a thumbs-up on each launch's GitHub Discussion; the week closes on Sunday at 23:59 UTC. Votes say how many people cared. The evidence next to each launch says whether the product calls Jev and who read the code. The two are never added together, and a rank is not this catalogue's recommendation.
How to launch a product ↗ · Each week's results as a feed (Atom)
Next week: 2026-W46
Launches accepted for next week, in no particular order. Their vote posts open on Monday 2026-11-09.
- ▲11-09
- ▲11-09
- ▲11-09
-
▲11-09
anydecisionmodel
Swift package for typed choices, scores and probabilities from Jev or MLX
- ▲11-09
No launch is on the board yet. A launch submitted by Thursday 23:59 UTC opens the next Monday: a maker can launch a product, and anyone can hunt one, such as those rising below.
New in the catalogue: 2026-W45
Products that entered the catalogue this week, in the order of what it records: who read the call site, then stars. An order, not a ranking. Their makers can launch them.
-
TypeSafe AI n8n node
TypeSafe's own n8n community node. Evaluate asks Jev typed questions about each item and adds the answers; Route sends each item to an output by one answer, with a fallback output for low-confidence choices and an uncertain one between two noul thresholds. AI Agents can use it as a tool.written by a model
-
Jevflake
dbt package (with a Terraform module) that lets Snowflake ask Jev about table rows: it sets up external access to api.typesafe.ai and creates SQL functions jev_noul, jev_choice, jev_score and jev_ask, plus dbt macros and tests so answers are stored once, reused and checked.written by a model
-
patrol-jev
Web app for Korean district-office street patrols: photos go in and a patrol log comes out as text or an .hwpx file. Claude or OpenAI turns each photo into text; Jev files it under one of four log sections, reads whether an address plate shows and judges whether two photos are the same spot.written by a model
-
Gut
Elixir DSL for LLM decisions: Gut.feel(subject, question, choices) returns one of your Elixir values, such as the team that should handle a ticket. Its Jev adapter asks the question as one Choice through ReqLLM; other adapters use any ReqLLM model, Ollama or your own code.written by a model
-
Jev Highlight Cutter
Agent skill and Python CLI that cuts long interviews, podcasts and streams into a highlight reel of a target length: Jev scores each transcript window on information value, hook, completeness, emotion and taste fit, plus yes/no gates, and FFmpeg cuts the chosen clips locally.written by a model
-
Jev steer or queue
Hook for Claude Code (and, more loosely, the Codex CLI) that asks Jev what to do with a message sent while the agent is busy: steer the running task now, queue it as a separate request for when the turn ends, or interrupt and stop the turn.written by a model
-
jev-remix-cutter
Agent skill and Python CLI that cuts a licensed remix from video: shot detection, beat or narration slots, Jev judging how well each shot fits each slot, constrained casting, local rendering and a review page, with originality and footage-licence reports.written by a model
-
jev-router (dominicrico)
Claude Code plugin that sends each prompt, step and subagent to the cheapest Claude model that can do the job (Haiku, Sonnet, Opus or Fable) and caps reasoning effort: Jev picks the model and effort as a Choice, and a live band above the prompt shows every pick.written by a model
-
jes
Open-source guardrails for AI agents: policies ask a decision model whether a prompt, retrieved page, skill, subagent, tool call, tool result or response carries prompt injection, a jailbreak, a data leak or a risky action, and block at your own threshold. Secrets and PII are redacted locally first.written by a model
-
jev-router (dirien)
Local pass-through proxy for Claude Code and the Codex CLI: on each message a person writes, Jev picks the model tier the work needs (a Choice), with optional yes/no guards for risky operations and routing claims; within a session it only moves up, to protect the prompt cache.written by a model
-
Jevaro
Python batching proxy that asks Jev the same questions about many states in one client request, either concurrently or packed several to a call, and streams the answers back in input order as Apache Arrow, with Python and JavaScript readers for pandas, Polars or DuckDB.written by a model
-
Jevboard
Windows tray tool that fixes homophone errors in Microsoft Bopomofo input before you press Enter: it reads the IME's candidates for each character position and asks Jev one Choice per position in a single request, then suggests corrections you apply with Tab.written by a model
-
jev-for-all
Connects Jev to OpenCode, Claude Code, Hermes and pi through one shared decision contract: Jev picks the model tier from the first prompt, which skill to load and which tools a step needs, and can start a lean session without skills or MCP servers. It does nothing when Jev is slow or unsure.written by a model
-
jevlint
Linter for code-review rules syntax cannot decide, each written as a plain-English yes/no question such as whether a secret reaches a log: Jev returns a probability per rule and chunk of code, reported like any linter's hits with files, lines, rule ids and exit codes. Rules are tested on fixtures.written by a model
-
Jev Mode
Plugin marketplace of Claude Code and Codex skills for building with Jev: Jevaluate has an LLM rate how well a project uses Jev, Jev Sources points an agent at TypeSafe's docs, a five-level study course, and the retired Jev Lens, whose check asks Jev nouls and choices about a finished state.written by a model
-
stefafafan/jev
Unofficial Unix command-line client for Jev, written in Go: pipe state on stdin, ask one noul, choice or score question per call and get the full answer as JSON, leaving thresholds to downstream tools such as jq. Works through TypeSafe, Cloudflare Workers AI or Vercel AI Gateway.written by a model
-
Jev Graph Search
Finds notes and passages in a local Obsidian vault or Logseq folder for an AI agent: it builds a shortlist on your machine, then asks Jev a relevance noul per candidate and ranks by it, returning the original passage with its source. Comes as a CLI and an agent skill.written by a model
-
Jev Guard
Guardrail for text an LLM app handles, whether user input, retrieved content, tool output or model output: Jev answers a noul for each of the ten OWASP LLM Top 10 risks, plus the attack type and a severity score, and a versioned local policy turns them into allow, review or block.written by a model
-
jev-apply
Job-application skill for CLI agents with a memory of your answers: Jev picks which saved answer fills each field of a Greenhouse, Ashby or Lever form, with an explicit none-of-these exit, and below a confidence threshold the field goes back to you instead of being guessed.written by a model
-
Jev-Mail
Gmail inbox classifier that runs on Google Apps Script and is managed from a CLI: for each thread Jev judges whether it needs action, whether it deserves a star and which configured category it belongs to, and the worker labels, stars or archives it. Ambiguous answers stay in the inbox.written by a model
-
jevwright
Browser tests for business flows written as the steps a user takes: Jev picks the control each step means on the live page, the path is recorded, and later runs replay it without a model. Your code, not the model, decides pass or fail, through declared requests, verify callbacks and invariants.written by a model
-
JEV Jornada
Workshop sales agent in LangGraph, in Portuguese, whose decisions (three guardrail nouls, a choice of which of seven queries answers the question, three nouls checking the reply is faithful) go to Jev, to an OpenAI or Anthropic LLM, or both, comparing accuracy, latency, tokens and cost.written by a model
-
jev-datafusion
Rust crate that registers Jev as SQL functions in Apache DataFusion: noul, choice, score and a multi-question ask run per row, so one query can score every ticket or route a year of email at warehouse scale. Works with TypeSafe, OpenRouter or any /v1/systemone endpoint.written by a model
-
jev-pilot
Claude Code plugin that asks Jev about each prompt before the turn: which reasoning effort it needs, which model and effort a subagent should get, whether to work directly, delegate or run in parallel, and which one skill applies. It raises effort once when tool calls keep failing.written by a model
-
jevelry
CLI, TUI and library keeping questions for Jev in small files called jevels; seventeen ship, for tickets, issues, pull requests, logs, alerts and failing tests. Each answer's certainty becomes act, mark for a person, or fall back, and every decision is logged so thresholds can be tuned on outcomes.written by a model
-
jevotron
Command-line field-level anomaly detection for CSV, JSON, text, SQLite and DuckDB: Jev labels each field of each entry NORMAL or ANOMALY, or with your own classes, given optional guidance and exemplars, and the output is a review queue sorted by confidence. SQLite caching reuses past answers.written by a model
-
pi-shift-router
Extension for the pi coding agent that routes each turn between a fast and a smart model tier, with failover on rate limits and exhausted quotas. Its judge can be a decision model such as Jev: one choice picks the tier and one noul decides whether to hand the task to subagents.written by a model
-
dsh-plugin-jev
Plugin for the DSH Web GUI that gives its agent a jev_ask tool for judgement calls and guards shell commands: before one runs, Jev scores how dangerous it is and whether its effects can be undone, on top of user-written blocked-command rules. Monitor mode by default; enforce blocks.written by a model
-
Jev Workflows
Codex and Claude Code plugin whose MCP tools put an agent's small judgements to Jev: classify a decision among the candidates it has, diagnose a failed command from its evidence, check whether evidence supports a completion claim and each acceptance criterion, or ask typed questions about one state.written by a model
-
jev-sim-use
Navigator for the iOS Simulator and Android devices built on sim-use: given a goal, each step is one Jev request asking which on-screen operation and target come next, whether that step finishes the goal and whether a tap is irreversible. When unsure it stops and leaves a session for your agent.written by a model
-
Jevflow
Claude Code plugin that holds the agent to a phased plan, each phase with a definition of done and a check. At each stop it runs them and asks Jev which phase the work is in, how far it got, and whether it is stuck, off goal or claiming done early; a tested policy says go on, next phase or done.written by a model
-
pi-jev-model-router
Extension for the pi coding agent that picks the model for each request: Jev names the kind of work, scores its complexity and the capability it deserves, and says whether it needs deep reasoning, while code composes those into a tier and weighs it against the spend caps and the budget used so far.written by a model
-
jev-demo
Three Spanish-language demos of one pattern, Jev decides and code acts: Jira ticket triage that applies, marks for review or escalates; a pre-flight pass that flags functions before Claude explains only those; and a PR router sending each pull request to rules, Haiku or Sonnet.written by a model
-
JevNext
Gets numbers out of Jev with ordinary Choice calls: each call picks a sub-interval or the next decimal digit, walking a multiway tree down to a finite-precision value, with no training or logits. Demonstrated on arithmetic, token billing and continuous game control such as CarRacing and LunarLander.written by a model
-
stingray
Claude Code Stop hook that catches turns that stop half-done: Jev judges whether the turn did nothing, declared an action its tool calls do not back, promised to watch work that is not running, or answered in the wrong language, and the hook blocks the stop with a fixed nudge.written by a model
-
xscout
Watches X for the updates that matter to one profile: Jev classifies each post's topic and speaker, checks exclusions, scores significance and relevance and picks the actions that apply, then deterministic rules add momentum and send big, trending or notable alerts to Slack with a drafting prompt.written by a model
-
adhere
Linter for rules a normal linter cannot check, written as Markdown with should and should-not examples. It asks Jev whether each file breaks each rule and reports those above a threshold with the part of the file Jev points at; adhere validate flags rules a linter could check or that contradict.written by a model
-
atoma
Platform that coordinates AI agents to turn a request into finished work. Jev takes its bounded decisions: which agent or stored recipe takes a task, with a noul on whether each candidate fits, and whether a plan or result can be approved; unsure answers go back to the run's models.written by a model
-
ContextGem
Python framework that extracts structured data from documents with LLMs and decision models. Given a decision model such as Jev, it asks a yes/no question per paragraph to find each aspect's clauses and turns Boolean, label and rating concepts into noul, choice and score questions.written by a model
-
fast-jev-compaction-opencode
OpenCode V2 plugin that compacts context by classifying instead of summarizing: it asks Jev, with noul questions, whether each completed tool call and its output still matter, keeps those and drops what can be re-read or re-run, and falls back to OpenCode's compaction if that fails or saves little.written by a model
-
Gallant
Go linter whose rules are questions for a decision model about a file's text, answered as noul, choice or score against a confidence threshold. By default it asks Jev through OpenCode Zen's free Jev model; other providers serve Liquid's, Cloudflare's or local models.written by a model
-
GoEventBus
High-performance Go event bus with an optional decision layer: when no deterministic rule or cached decision matches an event's state, Jev, reached through OpenRouter's Decisions API, picks which event type to run from a bounded candidate set, and execution stays inside the bus.written by a model
-
Harness
Command center that runs coding agents such as Claude Code and Codex side by side across machines. Its ⌘B router asks Jev 1.13 through OpenRouter's Decisions API, in one request of Choice questions, which session, project and agent a new task goes to.written by a model
-
Hindsight
Agent memory server that can hand two jobs to a decision model: re-ranking a recall pool with one Noul per candidate, optionally dropping unhelpful and duplicate memories, and settling easy consolidation cases with Choice questions before the LLM sees them.written by a model
-
http4k Connect: TypeSafe client
Kotlin client for TypeSafe's API in http4k Connect: a SystemOne action posts a state and Choice, Noul or Score questions to /v1/systemone with jev-latest as the default model, and a fake TypeSafe server answers the same requests in tests.written by a model
-
Jev Advisor Gateway
Local gateway between a coding agent (Codex, Claude Code, OpenCode, Kilo, Gemini CLI, Devin) and its model: Jev chooses the agent's next tool and answers noul questions about it, the session model writes arguments and the answer, and when Jev is unsure an optional read-only advisor adds a note.written by a model
-
laya-opencv
Local workbench and installer (Windows, Linux) for decision models. Inspection recipes turn OpenCV measurements into rule checks, then ask Choice, Noul and Score questions of local Laya, TypeSafe's Jev or aiask.me; disagreements and low-confidence answers are flagged for review.written by a model
-
Orchvia
Runs Claude Code and Codex agents as a team from your own application. Its optional routing layer asks a pluggable judge, Jev or plain rules, typed Choice, Noul and Score questions about which agent takes a request; below 0.85 confidence or a 0.2 margin the proposal waits for a person.written by a model
-
Parallax
DeepSeek Harness plugin for finding the arXiv papers on a topic in a local SQLite corpus: the agent recalls by keyword and combines sets, then Jev screens each paper against the inclusion criteria with a noul (kept at p ≥ 0.5) and sorts papers into topics with nouls or a choice.written by a model
-
SpotAsk
Mac menu bar app for quick questions with your own API key. Its External Ask can let a decision model such as Jev choose where a question goes (a web AI like ChatGPT or Perplexity, a local CLI agent, or a desktop app) with one choice question, then confirm inline or send above a threshold.written by a model
-
System One Computer Use
Prototype agent that drives a computer or an Android phone from screenshots: each step sends Jev the screen, the task and the history, Jev picks the next action (click, type, scroll, open a site or stop) and the text or site to use, and a Box question marks the element to click.written by a model
-
Yielded Agent
TypeScript runtime for durable agent threads on Effect and Effect AI. Its AutoModel picks a model profile when a thread is created, asking a decision model through Decision.classify which of the application's described models fits, and keeps that choice for later runs; the docs example uses Jev.written by a model
Rising in the catalogue: 2026-10-28 to 2026-11-04
Products whose GitHub stars grew fastest between the two latest weekly refreshes: at least 10 stars gained, in order of the gain as a share of the stars before. Stars are GitHub's count, a popularity signal, not a verdict, and nothing is awarded for rising.
-
★+35
jeveryword
Text extraction with Jev: field extraction, PII detection and exact quotes, built on TypeSafe's Jev.
-
★+12,379
agent-desktop
Desktop automation that reads the system accessibility tree and decides which button, menu or field to act on next.
-
★+26
anydecisionmodel
A Swift package for typed decisions from language models (probabilities, choices, and scores), with support for local MLX models and the TypeSafe Jev API.
-
★+53
fastbrowse
Browser agent: each page is indexed into candidate controls and Jev chooses the next one, so it cannot click what is not there; a Gemini Flash LLM plans and reads, and answers cite the page. Its maintainer recorded 99.2% on 46 tasks against 97.8% for hosted Browser Use, at under a tenth of the cost.
-
★+14
jev-guard
Security hook for coding agents: before each tool call Jev judges how much harm it could do, whether the user asked for it and whether it follows untrusted text, and code denies, asks or allows; afterwards Jev scans the result for prompt injection. For Claude Code, Codex, Cursor and more.
-
★+1,773
XERJ
Local-first search engine and MCP server in one Rust binary. Its optional `rerank` stage asks Jev one `noul` per top hit (does `document` answer the query in the state?) and re-sorts by the probabilities. The authors report FiQA nDCG@10 rising from 0.24 to 0.36, about $0.64 for 648 searches.
-
★+51
jev-docs-zh
Jev 模型（TypeSafe AI）官方使用文档的中文翻译 | Unofficial Chinese translation of the official Jev (TypeSafe AI) docs — https://docs.typesafe.ai
-
★+46
狗头军师 Jev Chat
Chat-reply assistant (desktop overlay, Android floating assistant, iOS keyboard) built on Jev Android. From the visible chat or a screenshot it asks Jev the literal question, the other person's intent, a danger level, whether to reply now and the best action, then has Jev pick among drafted replies.
-
★+10
Quicksilver
Claude Code skill that hands relevance calls to Jev: one noul per file, log line or ticket ('which of these files handle auth?'), asked in parallel and sorted by probability, so Claude gets a shortlist with confidences instead of reading everything.
-
★+15
openwhisper
Local speech-to-text, dictation, and meetings with Whisper and OpenAI API. Optional Windows x64 engines: Parakeet, Qwen3-ASR, Nemotron Streaming, and Moonshine.