User repository on GitHub
Calibrated tool selection and schema pruning for AI agents. Zero dependencies.
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Tool Prune
A zero-dependency tool selector that narrows a large pool of agent tools to the relevant few before the LLM call, running offline with TurboQuant or on TypeSafe Jev when a TYPESAFE_API_KEY is set.
- Stars
- 8
- Language
- JavaScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- hemanth
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- JavaScriptPython
- Platforms
- API
- Capabilities
- Open sourceRuns locally
Highlights
- Ships as npm and pip packages that default to the offline TurboQuant engine and route to TypeSafe Jev only when a key is present.
- filter() auto-selects candidates from the score drop-off or takes { k: 5 }, and the README claims prompt-token cuts of up to 92%.
- dispatch() runs a tool directly when confidence clears your threshold, described as under 160 ms with zero LLM generation cost.
- On Gorilla BFCL v3, the Jev engine is listed at 100.0% distractor top-1 at 189 ms; TurboQuant variants score 85-87%.
- TurboQuant uses turboquant-search (WASM SIMD) in JS or turbovec (Rust SIMD) in Python, with a built-in FWHT fallback.
Quickstart
bash
npm install tool-prune
# or
pip install tool-prune
# Optional: TypeSafe API key for cloud reasoning
export TYPESAFE_API_KEY="apikey_..."Watch out
Copy reviewed against the repo README on 2026-10-15; the pasted README was truncated mid-way through the 60-tool agent architecture benchmark table, so only its first row was visible. Benchmark numbers are the README's own, and the 92% token-cut claim is not tied to a named eval in the excerpt.
Questions about this entry
Short answers drawn from the linked project or page.
It prunes a large tool set down to the tools relevant to a prompt before the LLM call, so agents stop dumping every schema into context. prune() returns the winning match, filter() returns candidate tools, and dispatch() can run a tool directly when confidence clears your threshold.
No. prune() runs offline with the built-in TurboQuant engine by default. A TYPESAFE_API_KEY only routes decisions to the cloud TypeSafe System One (Jev) engine, which the README lists at 100.0% distractor top-1 on the multiple-tool BFCL v3 benchmark.
It detects the confidence drop-off in the score distribution and keeps the candidates above it, or you can pass a fixed count such as { k: 5 }. The README says this pruning cuts prompt tokens by up to 92% and eliminates context confusion.
Yes. router.dispatch() executes a tool directly when confidence clears your threshold, which the README describes as deterministic execution in under 160 ms with zero LLM generation cost, falling back to a callback you supply when confidence is too low.
On Gorilla BFCL v3, the TypeSafe engine is listed at 100.0% distractor top-1 and 100.0% 100-tool top-5 pruning at 189 ms, while TurboQuant variants score 85-87%. A separate 60-tool agent benchmark with Claude Haiku 4.5 starts with Tool-Prune Direct at 97.5% accuracy, 149 ms and 385 tokens per turn.
JavaScript via npm install tool-prune and Python via pip install tool-prune, with an identical API on both sides. TurboQuant uses turboquant-search (WASM SIMD) in JS or turbovec (Rust SIMD) in Python when installed, with a built-in FWHT fallback.
hemanth published it under the MIT licence; GitHub metadata shows the repository created on 17 September 2026 and pushed on 4 October 2026. The package advertises zero dependencies, and a live interactive playground is hosted on GitHub Pages.
More Jev resources like this
Open reimplementation of the Jev decision-model idea: masked-logit softmax turns any small local LLM or OpenAI-compatible API into a type-safe decision engine with raw probabilities. Also ships a WeChat/QQ/Telegram crush radar with a verdict and next-move advice.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
lookski/openjevPython
lookski/openjevPythonA policy engine for LLM decisions: declare routes, gates and actions once in TypeScript or Python and every decision comes back validated, explainable, replayable and audited. The model answers only declared questions — application code owns the branch.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
timmikeladze/jevlangJavaScript
timmikeladze/jevlangJavaScriptA fork and bridge that runs jkudish/jev-browser fully locally: a local Jev-style engine built on parallel-decisions replaces the hosted API, with Qwen2.5-1.5B-Instruct picking each browser step and a 0.5B model typing search queries, measured on an RTX 2060 SUPER.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
rorshopping/jev-browser-localPython
rorshopping/jev-browser-localPythonAsync, typed intent and tool routing library for Python agents: register your own callables with @router.intent and @router.tool, let Jev decide which capability handles an input, and let a confidence gate stop weak or invented decisions.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
huzeyfe07/jev-routePython
huzeyfe07/jev-routePythonA lightweight Python engine that mimics Jev's structured decisions on local LLMs via Ollama, vLLM or llama.cpp: choice, score and noul questions come back as JSON with probabilities and confidence, plus an OpenAI-compatible gateway.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
tapsin/jev-localPython
tapsin/jev-localPythonAn open-source Jev-based MCP router: it discovers, catalogs, classifies and deduplicates every MCP server and tool you have, then exposes each agent only the small, authorized subset it needs for the task. Routing decisions are System One Choice/Score/Noul questions, scored by a pluggable engine at the edge or hosted.
Jev Agents & MCP#community#catalog-import#logicrw
Community
amantux/mcp-routerPython
amantux/mcp-routerPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.