OPEN-SOURCE CATALOG

Tools

Browse community tools built with Jev. Ranked by GitHub stars, then forks, then name.

436 tools

  1. jev-ultrafast is a browser-use tool built for maximum speed, providing ultra-fast browser operations when execution latency is the priority.

    13,115807
  2. Laya is a non-autoregressive decision engine that answers typed questions (choice, score, noul) over text or JSON states in a single ~33 ms forward pass, trained with RL on strictly proper scoring rules for calibrated confidence, with a router that dispatches to one of three checkpoints across 100+ languages.

    5,763498
  3. A Claude Code plugin that replaces standard compaction summaries with Jev decision-making, scoring every tool call and result in a single fast request so stale items are dropped or truncated while retained content stays verbatim.

    5,523300
  4. SemIf adds semantic if-condition evaluation backed by open models running locally on a home NVIDIA RTX 3090, enabling meaning-based branching without depending on a hosted model service.

    2,624162
  5. A Monad trading bot that queries a TypeSafe Jev model each ~300 ms block for a buy/sell call on Kuru's MON-USDC book and posts a post-only limit order one tick inside the touch—earning spread on taker fills—with dry-run mode, sub-block two-RPC-round-trip hot loop, and an SSE dashboard server streaming per-block quotes, fills, and P&L.

    1,626306
  6. NanoJev is a 0.6B-parameter parallel decision model (Qwen3-0.6B backbone plus decision heads) that maps states, questions, and 2–255 candidates to complete probability distributions in a single forward pass—zero output-token decoding—with persistent serving and verified maze/Snake demos.

    1,592183
  7. A collection of agent skills for building with TypeSafe's System One API, giving coding agents reusable guidance and workflows that make it easier to integrate and work with the API correctly.

    1,31361
  8. kev is a local decision-model server — a LoRA adapter plus pointer readout head on a Qwen base (0.5B–8B) — that answers many typed questions (yes/no, multi-choice, ordinal score) about a document in one decode-free prefill pass, returning calibrated softmax probabilities through a TypeSafe System One–compatible API.

    1,24874
  9. Jevlike is a research starter for training a small one-pass model that scores N text options against a context via per-option attention queries and softmax, with train/eval/predict CLIs, a byte or frozen Hugging Face encoder, JSONL data, and Doom, chess, and Wikispeedia examples.

    1,10697
  10. A macOS computer-use tool that automates GUI interactions for roughly $0.0002 per step by OCR-ing the screen, classifying the next action with TypeSafe, and clicking.

    67747
  11. LocalJev is a Bun/TypeScript server exposing a Jev-compatible `POST /v1/systemone` decision API that runs typed questions on DiffusionGemma through any OpenAI-compatible endpoint (e.g. oMLX), prompting for JSON probabilities with validation, retries, normalization, and entropy-based confidence.

    64442
  12. Jev-cu is an installable Codex skill plus scripts that delegate Computer Use's next-step choice to Jev, which selects element, action, completion, and risk from text-only UI candidates while Codex executes, behind a dry-run-first local policy gate requiring human confirmation for sensitive operations.

    50350
  13. OpenAgents is a Rust monorepo home to Coder, a terminal coding agent that routes turns through typed Jev judgments and a model-agnostic generative door, runs a bounded shell loop, and talks to a self-hostable Nostr relay for account-free, inspectable job transport.

    45060
  14. Foreman is a Python asyncio supervisory runtime that pairs Codex coding workers with TypeSafe's Jev decision model, streaming factory events into ten probabilistic assessments so a deterministic policy can steer, retry, verify, finish, or escalate work without interrupting the agent.

    44332
  15. A TypeSafe Jev–based code-review tool that orchestrates staged model judgments (file profiling, evidence selection, mechanism classification, severity scoring, conditional reviewer routing) over a Git diff or full codebase to flag correctness, security, reliability, compatibility, and test-coverage risks, presenting results in a localhost-only dashboard.

    42925
  16. Simple Jev is a local Hugging Face/PyTorch server that turns shared context plus choice, rubric-score, or truth questions into JSON answers by scoring next-token logits, reusing the common prompt’s KV cache to avoid generated completions and repeated prefill.

    41744
  17. A self-hostable Cloudflare Workers app that turns plain-language requests into Search1API queries across web and vertical engines, then streams Jev-scored, URL-merged links and snippets with visible relevance, editable source/time filters, caching, and per-IP rate limits.

    34145
  18. An open-source TypeScript library that classifies each PDF page of a tax document into 261 IRS form IDs and 7 page kinds with calibrated confidence scores—via one call to TypeSafe's Jev decision model against a generated JSON form registry—at $0.001/page, 34× cheaper and 6× faster than an LLM pipeline.

    30930
  19. Eight agent-agnostic skills and a Hermes plugin that delegate per-turn decisions—model routing, memory filtering, compaction, skill selection, triage, GUI/browser actions—to TypeSafe's Jev decision model (~0.4 s, ~$0.00006), with redacted payloads and fail-open fallbacks.

    30825
  20. TypeSafe Mario is a Python harness that lets TypeSafe's Jev model play Super Mario Bros. by parsing emulator RAM into structured JSON describing Mario's motion, hazards, terrain, and timing, then translating the model's Choice/Noul/Score judgments into controller inputs, with a live telemetry dashboard and per-decision JSONL logging.

    30631
  21. A standalone mobile agent that drives a real Android phone through the Mobilerun API using TypeSafe's Jev model for per-step decisions—no ADB—bundling a live React streaming studio, a CLI with preview/execute modes, verified demos, JSONL execution traces, and request-level latency breakdowns for reproducible evaluation.

    29137
  22. jev-router is a loopback proxy that wraps the Claude Code and OpenAI Codex CLIs, sending each fresh user turn to a Jev routing call that picks a fast, balanced, strong, or long model tier while preserving native sessions, tools, auth, and manual model-picker overrides.

    27027
  23. pg-jev is a PostgreSQL extension that lets you filter, rank, and classify table rows with plain-language conditions via SQL functions (`jev`, `jev_prob`, `jev_score`, `jev_choice`) backed by TypeSafe's Jev model, which returns calibrated probabilities per row—using batched API requests, read-ahead streaming, and session caching—without indexes or embeddings.

    25514
  24. jev-align is an experimental CLI that iteratively improves TypeSafe Jev AI Functions by surfacing uncertain dataset rows for human labeling, running GEPA optimization with configurable reflection models and providers (TypeSafe, Vercel, Cloudflare), and resumable runs that can keep learning from captured production examples.

    24919
  25. FastAPI server exposing the TypeSafe/Jev evaluation API backed by Qwen3.6-35B-A3B on SGLang (Modal-deployable on B200), answering up to 64 Noul/Choice/Score questions per request via single-token logprob readouts over a shared radix-cached prefix.

    24028
  26. Open-source, Jev-compatible "System One" decision server: answers typed questions (yes/no, choice, score) about a state or images with calibrated probabilities in tens of milliseconds, plus OpenAI-compatible text generation, served by DiffusionGemma via vLLM.

    23421
  27. Skillbox is a self-hosted, versioned skills library for AI agents that combines Markdown editing, immutable revisions, scoped permissions, revocable client keys, MCP/CLI access, imports/backups, and optional model-backed recommendations to manage, distribute, and audit reusable agent skills.

    21918
  28. Drop-in replacement for typesafe_sdk's `system_one` evaluation API backed by OpenAI/Anthropic LLMs instead of TypeSafe, enabling cost/speed/intelligence comparisons with sync/async clients, structured outputs, probability normalization, retries, and usage/debug telemetry.

    21129
  29. Jev-browser drives a headless browser via MCP server, CLI, or library, with TypeSafe's Jev model choosing one action per step while code enforces budgets and stop gates, returning the final page, a step trace with confidences, errors, and a screenshot at sub-cent cost.

    20215
  30. Abide is a CLI that hooks Claude Code, Codex, and OpenCode to check every edit or turn against repository instruction rules via TypeSafe’s Jev decision model, catching violations linters miss, prompting in-turn repairs, and providing audit, calibration, cost, and reporting commands.

    19611
Original post
Start typing to find a use case.