Jev open source projects, MCP servers and integrations

TypeSafe Python SDK

The official Python client, with sync and async clients and built-in retries.

Jev

Calls api.typesafe.ai with your TypeSafe key.

TypeSafe JavaScript SDK

The official TypeScript/JavaScript client for Node.js 20+. Answer types are inferred from your questions.

Jev

Calls api.typesafe.ai with your TypeSafe key.

TypeSafe agent skill

A skill that teaches Claude Code, Codex and other coding agents how to write good Jev questions.

Jev

Documentation only. Your agent writes code that calls Jev.

TypeSafe n8n node

The official n8n community node (@typesafe-ai/n8n-nodes-typesafe-ai). Evaluate adds Jev's answers to each item; Route sends each item to an output based on one answer.

Jev

Calls TypeSafe's API with your TypeSafe key.

System One adapter (Python)

A drop-in replacement for the Python client that answers with an ordinary LLM API instead of Jev.

Jev

None. Useful for comparing Jev with an LLM through the same interface.

TypeSafe WorkflowEvals

TypeSafe's benchmark site comparing structured workflows (code rules plus choice, score and yes/no questions) with the same policy written as one prompt, across four example workflows. It reports accuracy, cost and time per case.

Jev

A results page, not a tool. Jev is one of the model configurations it compares.

Note: Vendor-run. Reference answers are the average of two LLMs, not human labels.

System One Connector (formerly typesafe-mcp)

A single Go binary (v0.4.12). One command registers it with Claude Code, Claude Desktop, Codex and Hermes. Named profiles switch between models, and a per-question min_confidence lets weak answers come back as uncertain.

Jev

TypeSafe or OpenRouter key. Profiles can also call Liquid AI's d1 or Kev through OpenRouter, or a self-hosted Laya or CLM server.

jev-mcp (jkudish)

Runs with npx -y @jkudish/jev-mcp (Node.js 22+). Twelve ready-made tools, such as verifying claims against sources, screening fetched pages, reranking and reviewing a diff.

Jev

TypeSafe, OpenRouter, Cloudflare Workers AI or Vercel AI Gateway key, or any Jev-compatible endpoint.

Note: The author calls it early software; expect rough edges.

JevMCP

A Python server that reads files by glob pattern and asks Jev about each one, so your agent doesn't load every file into its context.

Jev

TypeSafe or OpenRouter key.

Note: Early release (0.1.0) with no updates since 22 September 2026; treat it as dormant.

jev-mcp (codaaiteam)

Runs with npx. Five tools: classify, score, check, gate a risky action, and a general decide tool.

Jev

Official TypeSafe key goes to api.typesafe.ai.

Note: Keys starting jv_live_ are sent to a third-party paid gateway (jevtypesafeai.com), not to TypeSafe. It charges $0.25–0.42 per 1M input tokens, 6–10× TypeSafe's $0.042.

openjev-mcp

A Node.js server with one general tool and three single-question shortcuts.

Jev

Calls the third-party OpenJEV API (api.openjev.sh), not TypeSafe directly.

Note: Your data passes through a third-party service. OpenJEV says its inference is paid for by trading fees on a $JEV crypto token; it is not affiliated with TypeSafe.

Vercel AI Gateway

Model id typesafe-ai/jev, called with the AI SDK's experimental_decide. No markup on the token price.

Jev

Proxies to TypeSafe. Per-request zero data retention option.

OpenRouter Decisions API

POST /api/alpha/decisions with model typesafe/jev-1.13. Each response reports its cost.

Jev

Proxies to TypeSafe, billed to your OpenRouter account.

Note: The endpoint is still on an alpha path and may move.

Cloudflare Workers AI

env.AI.run('typesafe/jev', { state, questions }) from a Worker.

Jev

Proxies to TypeSafe.

LiteLLM pass-through

Forwards /typesafe/* on your LiteLLM proxy, with cost tracking and logging.

Jev

Uses your TypeSafe key from the proxy environment.

Spring AI

Structured judgments from Java and Spring Boot applications.

Jev

Calls TypeSafe's API.

LangChain

TypeSafeClassifier, a LangChain Runnable, from pip install langchain-typesafe (Python) or npm install @langchain/typesafe (JavaScript). The Python package adds experimental middleware for routing agent runs between models and blocking risky tool calls.

Jev

Calls api.typesafe.ai with TYPESAFE_API_KEY; TYPESAFE_BASE_URL points it at a compatible gateway.

Note: Both packages are early (Python 0.0.1a3, JavaScript 0.0.2).

TanStack AI

The @tanstack/ai-typesafe adapter for TanStack AI's decide(). It covers decisions only, with no chat or embeddings.

Jev

Plain fetch calls to POST /v1/systemone on api.typesafe.ai with TYPESAFE_API_KEY; baseURL can point at a gateway.

Netlify AI Gateway

Use TypeSafe's own @typesafe-ai/sdk inside Netlify Functions with no API key or base URL to configure.

Jev

Proxies to TypeSafe. Usage is billed to your Netlify credits.

Zuplo AI Gateway

Model ids jev/jev-latest, jev/jev-1.13.0 and jev/jev-preview on your gateway app's /v1/systemone, with per-app keys, token budgets and a cost header on every response.

Jev

Uses your TypeSafe key, stored in the Zuplo provider settings.

Note: Zuplo's DLP, firewall and prompt-injection guardrails block Jev requests by default because they can't inspect them, and requests never fall back to another model.

Jev Router (OpenRouter)

A chat model id, typesafe/jev-router, that uses Jev to pick a model and reasoning effort for each chat request. Call it through OpenRouter's Chat Completions API.

Jev

Jev makes the routing decision; another model writes the reply.

Note: OpenRouter lists no fixed price: you pay for whichever model it routes to.

Laya

A 421M-parameter encoder (plus a 322M multilingual one) answering choice, score and yes/no questions. pip install laya (v0.4.2).

Jev

None. Its server speaks the same /v1/systemone format, so Jev clients can point at it.

Note: Reads 512–1,024 tokens per question by default (the multilingual model up to 8,192 with max_len=8192), less than Jev's 32k.

Open-Jev

LoRA adapters and a decision head on top of Qwen models (2B, 9B, 27B).

Jev

None. An independent re-implementation of the idea.

Note: Base weights are downloaded separately; it does not reproduce TypeSafe's training method.

CLM-v0.1-8B

An 8B open System One model whose server implements the same /v1/systemone endpoint.

Jev

None. Works with Jev MCP servers that accept a custom base URL.

Clef and Clef-Flash (Cloudflare)

Cloudflare's 27B and 9B decision models, fine-tuned from Qwen. They read text, JSON and images; Clef also reads video.

Jev

None. Cloudflare says the API is fully compatible with Jev; Clef is also hosted on Workers AI at $0.24 per 1M input tokens.

Note: Self-hosting needs a data-centre GPU. Accuracy figures are Cloudflare's own.

pplx-decider (Perplexity)

Perplexity's 27B decision model (v1.1, 5 October 2026), fine-tuned from Qwen. Accepts images.

Jev

None. An independent model answering the same question types.

Note: Perplexity reports a higher Decision Index score than Jev; that figure is its own. Text and images must fit in 8,192 tokens.

Kev

Jared Palmer's family of decision models in 0.8B, 4B, 9B and 27B sizes, built on Qwen. The 0.8B runs on a laptop, and you can train your own.

Jev

None. Its API matches TypeSafe's, so the TypeSafe Python SDK works against a Kev server; Kev-4B is also on OpenRouter.

Strands Decider (AWS Strands Labs)

A small decision model from the Strands Agents team, with a 2B version built on Qwen3.5 and Gemma-based sizes. pip install strands-decider gives a CLI that runs it on a laptop CPU, Apple silicon or a CUDA GPU.

Jev

None. An independent model answering choice, score and yes/no questions.

Note: Reads up to 4,096 tokens. Calibration figures are the authors' own.

Mobile UI automation

An Android automation prototype that selects from visible UI actions.

Jev

Selects a UI action.

Note: Requires a device or cloud phone; reaching checkout does not mean an order was completed.

Layered code review

Groups code changes by the type and severity of review they may need.

Jev

Assesses risk, candidate issues, and severity.

Note: Not proof of an automated audit; it does not execute user code.

Semantic Postgres filtering

Uses Jev to decide whether database records match a text condition before filtering or sorting them.

Jev

Decides whether a record matches a text condition.

Note: The example conditions are not standard SQL; data may be sent to an external service.

Music theory choices

Chooses from defined notes, chords, or other music-theory options for playback.

Jev

Evaluates or selects music-theory candidates.

Note: Jev is not an audio generation model; timbre comes from code or supplied assets.

Jevthoven composition

Chooses instrument, chord, and rhythm options that application code turns into MIDI events.

Jev

Selects instrument, chord, or rhythm candidates.

Note: This combines bounded musical building blocks; it is not open-ended composition.

Snake decisions

A small game demo where code executes Jev-selected moves.

Jev

Selects a legal movement direction.

Note: A legal move is not necessarily optimal; prerecorded replays should not be presented as live.

Stealth guard NPC

Chooses a game guard's next action from the state and available actions supplied by the game.

Jev

Assesses suspicion, targets, and actions.

Note: The decision input is structured state, not an unrestricted understanding of the game screen.

Jev plays chess

A documented chess experiment comparing typed move selection.

Jev

Selects the next move from legal candidates.

Note: The author's experiment is not this site's chess benchmark and should not be extrapolated into a ranking.

Automatic slide advance

An author demo that decides when a presentation should move forward.

Jev

Evaluates semantic progress through the presentation.

Note: Speech recognition comes from another system; the initial site adaptation uses text only.

Semantic Unix pipes

Filters or scores records in a command-line workflow using Jev's answers.

Jev

Filters, classifies, or scores supplied records.

Note: This site does not execute arbitrary shell commands.

Knowledge graph navigation

Chooses which connected item to follow next in a knowledge graph.

Jev

Selects the next relationship to follow.

Note: This is not a verified global shortest-path solver.

Agent completion checks

A public project that evaluates whether an agent has evidence of completion.

Jev

Provides a semantic signal about a completion claim.

Note: Jev cannot replace actually running the tests.

JevBird trajectories

Chooses a movement path from options defined by the game code.

Jev

Selects from candidate movement strategies.

Note: This is not direct, continuous control of the physics simulation.

Sponsor segment detection

A project that evaluates transcript segments for video skipping.

Jev

Classifies transcript segments or assigns probabilities.

Note: Classifying a transcript is not the same as directly understanding video.

json-render composition

Chooses from predefined interface components that a renderer displays.

Jev

Selects a bounded component or specification.

Note: The related APIs are experimental and should not be treated as stable dependencies.

Instinct UI generator

Chooses predefined fields, components, and styles for an interface.

Jev

Selects component, field, and style candidates.

Note: Jev does not directly generate arbitrary source code in this project.

Wikipedia pathfinding

Chooses which available Wikipedia link to follow next during a path search.

Jev

Evaluates candidate links.

Note: This site's sandbox does not crawl arbitrary pages or claim a real-time shortest path.

Audio beep automation

A project that identifies where supplied recordings should be censored.

Jev

Decides whether a word should be muted under the supplied rules.

Note: Jev does not perform transcription or audio editing; real-time operation has not been verified.