Agents: read the raw Markdown of this page, or start at llms.txt.
Community SDKs, libraries and open replicas
TL;DR TypeSafe ships SDKs for Python and JavaScript only (Python SDK: install, clients, system_one(), JavaScript/TypeScript SDK quick contract: install, client, choice/score/noul, answers, errors (the page a builder reads)); Java goes through Spring AI (Platforms and gateways: Zapier, LangChain, Spring AI, Cloudflare, Netlify, Vercel, OpenRouter, Fly.io, Pydantic AI and other hosted routes to Jev). Everything below is unofficial community code (SDKs, libraries, SQL functions), days old (★ and push as captured 2026-09-23 to 25): pick the most-tested client for your language, check its name is the one you mean, and keep the key server-side. Open replicas speak
/v1/systemonebut are not Jev and lack its calibration. Split from Tools and integrations: MCP servers, skills and hooks, plugins, CLIs and dev tools on 2026-09-25 (nothing dropped); command-line tools, hooks and MCP servers stay there.
Licence MIT unless stated. Numbers are the author's. Pattern IDs: Decision patterns from the community (with fit verdicts).
Community SDKs by language
Names that look official are not. RubyGems typesafe-sdk and crates.io typesafe-ai are community packages; TypeSafe's own packages are PyPI typesafe-sdk and npm @typesafe-ai/sdk (typesafe-ai GitHub organisation and repos).
| Language | Project (owner) | Install | ★ · push | Note |
|---|---|---|---|---|
| Go | typesafe-sdk-go (Radix ILS) | go get |
0 · 2026-09-18 | Python SDK 0.6.0's wire format; TYPESAFE_* names verified |
| Go | jev-go (Gaurav-Gosain) | go get github.com/Gaurav-Gosain/jev-go |
5 · 2026-09-16 | helpers YesNo, OneOf, Levels; base of jev-sec-bench (Head-to-head: Jev against other models and methods) |
| Go | jev-go (Stumble) | go install github.com/stumble/jev-go/cmd/jev@latest |
5 · 2026-09-18 | stdlib only; direct (TYPESAFE_API_KEY) or Vercel Gateway (AI_GATEWAY_API_KEY); CLI refuses keys as flags; says "community-maintained", its agent skill defers to the official one |
| PHP | laravel-typesafe-jev (Butochnikov) | composer require |
3 · 2026-09-17 | Laravel facade, async, Jev::fake(); on his typesafe-sdk-php |
| R | jevr | pak::pak() |
4 · 2026-09-20 · NOASSERTION | jev_ask(); TypeSafe or OpenRouter |
| Elixir | jev (dannote) | Hex {:jev, "~> 0.1"}; Elixir 1.18, OTP 27 |
33 · 2026-09-22 | Jev as a peer process: answers arrive as messages you pattern-match; handlers test without network. P06 |
| Ruby | typesafe-sdk (joshmn) | gem install typesafe-sdk |
8 · 2026-09-24 | mirrors the Python SDK's answer shapes. Gem name equals TypeSafe's PyPI name and the README has no "unofficial" line |
| Ruby | ruby_llm-typesafe (kieranklaassen) | gem 'ruby_llm-typesafe', RubyLLM ≥ 2.0.0.rc3 |
19 · 2026-09-16 | RubyLLM provider, structured output only: chat without a schema, streaming and tools raise before any request |
| Ruby | typesafe-ai-rails (GenieRobot) | gem, on joshmn's gem | 4 · 2026-09-16 | fail-closed act! refuses side effects without a stored confidence policy. Its example model: "jev-1.12" with extra_body: {beam_width: 4}: neither in Models, aliases, pricing, rate limits, context or the API, unverified |
| Rust | typesafe-ai-rs (gilljon) | cargo add typesafe-ai-rs |
6 · 2026-09-17 | async and blocking, cancellation, full response metadata; targets SDK 0.6.0 (Python is 0.7.1) |
| Rust | typesafe-ai (Twister915) | crate typesafe-ai |
13 · 2026-09-16 · Apache-2.0 | retries observable as an iterator/stream; example CLIs for PR questions and tsg semantic search. Crate takes the company's name |
| Rust | typesafe-rs, s1-rs (AbdelStark) | — | — | names only (not captured) |
| Scala | zio-typesafe-ai (jamesward) | libraryDependencies += "com.jamesward" %% "zio-typesafe-ai" % "<version>" |
5 · 2026-09-23 · Apache-2.0 | ZIO client; out-of-range requests can't be built (its "1–255 options": docs state the 255 max only, verified); opt-in logging never logs headers or the bearer token |
| .NET | typesafeai-dotnet-sdk (saibimajdi) | dotnet add package TypeSafeAI.Sdk |
9 · 2026-09-23 | forward-compatible (RawJson, UnknownAnswer), trimming/AOT-clean; states "not affiliated" |
| Kotlin | jev4k (Paul Brose) | Maven com.pambrose:jev4k |
6 · 2026-09-20 · Apache-2.0 | Kotlin DSL, enum-valued Choice; docs jev4k.com. Also kojev (name only) |
| Swift | swift-typesafe (ainame) | SwiftPM from: "0.7.2" |
15 · 2026-09-23 | follows Python SDK 0.7.1; Swift 6.4 on Apple's experimental HTTPClient; Linux via AsyncHTTPClient |
| Python (async) | jevclient (AboveColin) | pip install jevclient |
3 · 2026-09-21 | checks options and 2–10 levels locally (255-option max and 2–10 levels verified); latency table below |
jevclient's cost of batching (author, cold connection from the Netherlands, 2026-09-17): one call barely slows as questions grow, but every question is billed input (~38 tokens each for short ones; output free, verified).
| Questions | Input tokens | Latency |
|---|---|---|
| 3 | 521 | 712 ms |
| 25 | 1,242 | 690 ms |
| 100 | 4,017 | 714 ms |
| 400 | 15,417 | 1,336 ms |
Warm connection: 250–580 ms over 16 calls. Consistent with Speculative fan-out.
Libraries and database functions
Code you import, or SQL you call. Command-line tools, hooks and MCP servers: Tools and integrations: MCP servers, skills and hooks, plugins, CLIs and dev tools.
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| jev-harness (Antonio Coppe) | TypeScript: policy, confidence gate, shadow mode, recipes, offline eval CLI | npm install jev-harness |
15 · 2026-09-25 | 24-row filter 1.3 s vs 48.9 s with claude -p, one machine. Not Astro-Han/jev-harness (Builds: coding agents, harnesses and orchestration). P04, P06 |
| hono-jev-router (yusukebe) | Hono router: routes described in words, one Noul each, first at or above threshold (default 0.5) wins |
npm i hono-jev-router |
50 · 2026-09-18 | not an auth boundary (the author says so: the requester steers the input); auth headers redacted before sending. P13 |
| jev-tree (reachjalil) | recursive Choice past the 255 cap (verified) |
npm install jev-tree |
6 · 2026-09-18 | Cookbook: Hierarchical classification |
| JevGuard (Seb4Ez) | cache, UNRESOLVED_OR_OTHER option, AMBIGUOUS_STATE flag |
pip install jevguard-core |
0 · 2026-09-22 | its "official" MCP is not TypeSafe's |
| jevcal (abhixhek) | thresholds per accuracy target; CI drift | pip from git (PyPI jevcal is a different project, aakgna/jevcal, per our 2026-09-24 registry check) |
10 · 2026-09-18 | MCA benchmark-ban claim contradicts docs since 2026-09-19 (Legal: MCA, DPA, privacy, data retention). P06 |
| hunch (steven-shoemaker) | pandas/list verbs classify, where, extract, rank, check; dedupes rows, caches answers; uncertain=(low, high) sends borderline rows to review |
pip install hunch-jev (PyPI hunch is unrelated) |
2 · 2026-09-22 | not Kelbie/hunch; its "frontier-level intelligence" line is marketing. P18, P42 |
| Advocaat (pithings) | tagged templates ask.if, ask.choice, ask.score with typed labels; agent skill |
npx nypm i advocaat |
92 · 2026-09-18 | TypeSafe or Vercel Gateway |
| discern (doeixd) | pattern matching and procedures over Effect's Decision/DecisionModel, "maybe" a branch you handle |
npm install @doeixd/discern effect@rc |
13 · 2026-09-22 | provider-neutral; Jev via @effect/ai-typesafe. P06 |
| jevals (openlayer-ai; Reddit write-up) | agent evals and guardrails as typed questions, all checks for a trace in one request; same class offline, as monitor, as gate | pip install jevals |
85 · 2026-09-24 | Jev via TypeSafe, Vercel or OpenRouter, or Kev/Laya locally. Not jevals.com (Head-to-head: Jev against other models and methods). P15, P16 |
| pytest-jev (allebee) | semantic assertions, batched Nouls per test run | pip install pytest-jev |
3 · 2026-09-21 | 12 designed tests × 3 runs: same verdicts as Haiku 4.5 and Sonnet 5, 5.3 s vs 15.4 / 27.1 s. P15 |
| llama-index-jev (WiktorB2004) | LlamaIndex JevRerank (0–3 relevance Score) and Jev selectors |
pip install llama-index-postprocessor-jev / llama-index-selectors-jev |
7 · 2026-09-25 | BEIR NFCorpus nDCG@5 0.340 → 0.396, ~$0.0003/query (Head-to-head: Jev against other models and methods); rerank fails open, select fails closed; not affiliated with LlamaIndex. P17, P02 |
| langchain-skill-router (deyna256) | loads only the skills a turn needs; Jev as judge | pip install "langchain-skill-router[jev]" |
4 · 2026-09-24 | 236 skills: 4.4× less input, accuracy +1.5 pts (p = 0.63, not significant); a wrong skill: 72% vs 92% correct (Head-to-head: Jev against other models and methods). P08 |
Database functions (your rows go to TypeSafe; batch where the tool allows).
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| pg_typesafe (Giulio Piccolo) | Postgres functions, batched *_many |
make install |
82 · 2026-09-20 | 38 texts: 23 s vs 0.86 s batched. P17 |
| jevql (kylemclaren, post) | jev() predicates in vanilla Postgres, rewritten client-side (no extension); CLI, HTTP and MCP node, Go/TS/Python SDKs |
brew install kylemclaren/tap/jevql |
13 · 2026-09-19 | never run serve --insecure open: it spends your key for anyone. P17 |
| sqlite3-jev (mattn) | SQLite functions jev_noul, jev_choice, jev_score, *_json, jev_raw |
build, .load ./jev |
3 · 2026-09-18 | first SQLite route; can target a local "tensai serve" instead. P17 |
| duckdb-jev (Colliber) | answers as ENUM, DOUBLE, STRUCT |
DuckDB extension | 23 · 2026-09-18 | P17, P18 |
Elsewhere: pg-jev → Repos: data, documents, judging, real-time, markets, business apps, replicas.
Open replicas and Jev-compatible servers
Trust note. None is TypeSafe or runs Jev. /v1/systemone compatibility is the wire shape, not the calibration Confidence-gated routing needs; several compute their own confidence. "No Jev outputs used" claims cannot be checked from outside. Re-measure on your labels (Measurements, access routes and open replicas).
| Tool | What | Install | ★ · push | Note |
|---|---|---|---|---|
| Kev (Jared Palmer) | 0.8B/4B/9B/27B on Qwen bases; serves /v1/systemone (+ /v1/systemone/permute) |
HF weights, uv sync |
6875 · 2026-09-25 · Apache-2.0 | own table: Kev-27B 0.848 vs Jev 0.857 on new sources; TypeSafe's 102 public eval questions: Kev-9B 0.809 vs published Jev answers 0.883; 13 long documents exceed its 8,192-token serving context; trails on knowledge (MMLU Kev-27B 0.84, Jev 0.90). States "No Jev outputs were used for training": unverified |
| decider (Mapika) | Qwen3.5 fine-tunes (2B, 4B, 35B-A3B), local Qwen3.5-27B teacher | pip install decider-ai |
394 · 2026-09-25 · Apache-2.0 | JevBench public hard tier Jev 0.730 vs decider-35b-a3b 0.676; publishes that its 2026-09-24 models failed their pre-registered release rules; "Nothing was distilled from Jev": unverified |
| JEVfire (kikoncuo) | Jev-inspired parallel decisions on vLLM (CUDA) and WebGPU | clone | 67 · 2026-09-18 | 28 decisions in 497 ms, 10.3× faster than generating the same constrained JSON on the same 27B model; Mario 71 ms/action in a browser (Qwen3.5 0.8B, M4 Max); credits the RLCD checkpoint (Measurements, access routes and open replicas) |
| PlayJev (OmniJev) | Qwen3.5-0.8B-Base playing ten browser games from raw pixels | HF | 36 · 2026-09-24 · Apache-2.0 | 43 ms per move on an H200; 0.57 of its game teachers (trained with DAgger). Jev itself is text-only (Models, aliases, pricing, rate limits, context) |
| Laya (Nandakishor M) | 3 encoder checkpoints + router; laya-serve |
pip install laya |
18593 · 2026-09-23 · Apache-2.0 | 32.8 ms on a T4; Jev figures quoted, not measured; Jev leads Banking77 0.870 vs 0.425; Khmer 0.000 at 0.952 confidence |
| Von (wfzyx) | 395M encoder | pip install von-sdk / npm install von-sdk (plain von on PyPI and npm is unrelated) |
526 · 2026-09-23 · Apache-2.0 | own table: Jev 96.6% vs Von 72.0% on 49 tasks |
| AnyJev (Nokia Applied Research) | any HF LLM, debiased, optional head on 100–300 labels | pip install "anyjev[hf]" |
267 · 2026-09-23 · Apache-2.0 | order flips 0.230 → 0.073 (Qwen3-8B) |
| Simple Jev (Eugene Cheah) | logits server, free demo API | pip from repo | 491 · 2026-09-21 · Apache-2.0 | own contract |
| jev-rs (Yijun Yu) | Rust over llama-server; serve, MCP, eval | cargo install jev-rs |
5 · 2026-09-22 · Apache-2.0 | SDKs reach it via TYPESAFE_BASE_URL |
| hearim (ziozzang) | Go gateway over Ollama, vLLM, SGLang | go build |
5 · 2026-09-22 · NOASSERTION | entropy-based confidence |
| openJev-verdict-2.0 (Heman10x) | 149.6M ModernBERT, confidence head, WebGPU | Hugging Face | 273 · 2026-09-20 · NOASSERTION | claims 77.10% vs Jev 72.70%, Jev's figure copied from Laya, not measured |
| visual-jev | image decisions, Qwen3-VL | uv | 2 · 2026-09-20 · no licence | Jev is text-only (Models, aliases, pricing, rate limits, context) |
| JevTuner | Brier-loss training recipe | clone | 12 · 2026-09-21 · no licence | not TypeSafe's method |
| Recipe (@theanandprasad) | softmax over yes/no logits | — | post 2026-09-22 | 21/22 on Llama 3.3 70B; raw logits flip with option order (AnyJev) |
Names only: LitJev (logit reads on stock Qwen), jevmlx (restricted softmax on MLX), jev-visual (teaching repo, not visual-jev). Elsewhere: SemIf, NanoJev, jevlike, openjev-sglang, mini-jev → Repos: data, documents, judging, real-time, markets, business apps, replicas.
Warning. Tools that train a local model on Jev's answers breach MCA §2.3(b), which forbids distillation and training a model to imitate the Services' output (Legal: MCA, DPA, privacy, data retention); e.g. stuntd's paid-upstream mode, which learns from TypeSafe's answers and then takes the decision over. Not listed as a tool.
Related
- Tools and integrations: MCP servers, skills and hooks, plugins, CLIs and dev tools — MCP servers, skills, hooks, plugins, CLIs
- Platforms and gateways: Zapier, LangChain, Spring AI, Cloudflare, Netlify, Vercel, OpenRouter, Fly.io, Pydantic AI and other hosted routes to Jev — vendor-shipped frameworks and hosted routes
- Measurements, access routes and open replicas — replica trust; TYPESAFE_* environment variables across SDKs —
TYPESAFE_*names a client should read
Sources
Links inline; raw captures (2026-09-23 to 25) in frontmatter.