Why not just ask the model I already have in the same agent?

The README's answer is that for a one-off question you probably should. The difference shows up in loops: an LLM's stated confidence is not calibrated, a call costs roughly 100 times more, and it takes seconds rather than milliseconds. Per item across a batch, that gap compounds — which is the case the server is aimed at, alongside triage, routing, tie-breaking and claim-versus-evidence checks.

Where do the cost and latency numbers come from?

They are the project's own measurements, labelled indicative only and priced at $42 per 1B input tokens. A choice with six rich options plus context runs about 0.6-0.7 s on roughly 1.7k tokens; a noul or verify with short context runs 0.4-0.6 s on 0.3-0.5k tokens; classify over 100 items takes about a minute sequentially. Re-measure on your own payloads before budgeting.

Can I keep the API key out of my config file?

Yes. Skip the env block entirely and call the setup tool once from your agent: it verifies the key live and stores it with 0600 permissions under the config directory. TYPESAFE_API_KEY in the environment still takes precedence over the stored file, so you can override per shell. JEVMCP_BASE_URL also overrides the endpoint and JEVMCP_MODEL the model name, which defaults to jev-latest.

Which Jev MCP server should I actually pick?

There are several. The README points at jkudish/jev-mcp, a Node/npm server offering ten workflow-shaped tools, and notes a Go server also exists. This one's case is the Python/uvx side with the official question types kept faithful, a one-command installer covering several clients, bilingual documentation and MIT. If you would rather have workflow-shaped tools than the raw primitives, that is the trade to weigh.

How do I install it into my MCP client?

Run uvx jev-mcp-server install claude-code, or substitute pi, cursor, opencode or codex, then restart the client if it is running. A plain pip install jev-mcp-server works too, and the installer subcommand is the same either way.

What are the limits on options, levels and batch size?

choice picks 1 of 2 to 100 options and returns probabilities over all of them so near-ties stay visible; score grades on an ordered 2 to 8 level rubric; and classify runs up to 100 items against one shared category set.

How do I cache identical calls?

Set JEVMCP_CACHE=1 and identical question payloads are answered from disk at zero cost, with usage.cached true. classify also deduplicates repeated items on its own, which matters when the same queue is labeled more than once.

CommunityJev Practices & Patterns

Jev MCP Server

Python MCP server exposing Jev's three official question types plus compare, verify and batch classify, with an installer subcommand that writes the client config and a measured cost per call.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Author: wangkuangkuang
Last verified: Oct 2026

wangkuangkuang/jev-mcp-server

User repository on GitHub

View on GitHub

MCP server for Jev (TypeSafe System One): the three official question types — choice, score, noul — plus batch classify. Calibrated probabilities, ~0.5s, <$0.001/call.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

MCP server for Jev, TypeSafe's System One model. It exposes the three official question types (choice, score, noul) plus compare, verify, batch classify, and a one-command installer that writes your client config for you.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Seven tools: choice, score, noul, compare, verify, classify and setup, with choice returning probabilities over every option so near-ties stay visible.
Measured per call: choice around 0.6-0.7 s and $0.00007, noul or verify 0.4-0.6 s and $0.00002, classify 100 items about a minute and $0.007.
uvx jev-mcp-server install claude-code writes the config for claude-code, pi, cursor, opencode or codex; plain pip install works too.
JEVMCP_CACHE=1 answers identical payloads from disk at zero cost, and classify deduplicates repeated items on its own.

Watch out

This is not the only Jev MCP server — the README names a Node/npm one with ten workflow-shaped tools and notes a Go server exists — so the argument for this one is the Python/uvx side, faithful official question types, the one-command installer, bilingual docs and MIT. It is also explicit about poor fit: open-ended reasoning and long-context work, and anything needing an explanation attached, since Jev returns numbers and any reason your assistant adds is its own reading of them. The cost and latency figures are labelled indicative and priced at $42 per 1B input tokens, so treat them as order of magnitude. You need a TypeSafe key; there is no local model and no helper LLM. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-rag-benchmark

Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

erendikmenn/jev-rag-benchmarkPython

jev_project_context

Research-memory skill that keeps experiment claims traceable with ten operations, provenance gates, and optional batched Jev audits over memory files.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

poiuyjie/jev_project_contextPython

jevmory

Local memory layer that stores verbatim agent quotes in per-project SQLite and grades them with Jev into stale, wrong, unsupported, or keep verdicts.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

romiluz13/jevmoryPython

local-jev

Local offline System One server speaking the hosted /v1/systemone wire format, so existing SDK code runs unchanged against small on-machine models.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

amithgc/local-jevPython

laya-browser-agent

Local harness that runs the open Laya decision model on your own machine and wires it into browser agents through a TypeSafe-compatible /v1/systemone dialect.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

ChenneyZhuang/laya-browser-agentPython

daf-jev

Python client and decision toolkit for the Jev System One endpoint with noul, choice, and score primitives, batched asks, composition helpers, concurrent evaluation, and an MCP server.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

docxology/daf-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Designing the State: What to Include and What to Drop

Oct 12, 2026

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.

Jev Practices & Patterns#state#design#primitives

Read article

Noul, End to End: One Probability and a Dozen Thresholds

Oct 12, 2026

A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.

Jev Practices & Patterns#noul#primitives#confidence

Read article

Scoring Without Measuring: Bands, Rubrics, and Composite Triage

Oct 12, 2026

A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.

Jev Practices & Patterns#score#primitives#scoring

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai