Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

tenbin

MCP server and agent skill for designing Jev judgments: question linting, batch evaluation, threshold calibration and reranking.

MCP serverbetaMIT
View on GitHub ↗

Overview

Tenbin helps coding agents use TypeSafe's Jev at design time. It has three parts: documentation mapping the official docs to workflows and use cases, a stdio MCP server, and an agent skill with procedures, templates and an offline linter. Asked for a judgment the code needs, the agent returns a measured design rather than a prompt: atomic Choice, Score and Noul questions, linted for problems like compound questions, counting or date math; raw answers on sample inputs; accuracy per confidence band on labelled rows; and thresholds written to one file with the measurement date.

The MCP server also provides reranking from one Noul per pair or a tournament of Choices, beam search down a category taxonomy, a cost guard and integration-code generation. Generation and linting need no TypeSafe key; paid evaluation runs only when requested. Production code calls the TypeSafe SDK directly and never depends on the MCP server. The README includes Japanese usage examples and can be registered with Claude Code.

What the decision model decides

TypeQuestionAnswers
Yes / noDoes the customer want a refund?—
ChoiceWhat caused the damage?shipping_damage · other · ...
ScoreWhat is the tone (4 situation-based levels)?—

How it works

  1. Ask the agent for a judgment the project needs.
  2. Decompose it into atomic questions and lint them offline.
  3. Optionally evaluate on labelled rows to measure accuracy per confidence band.
  4. Write thresholds and production SDK code that calls Jev directly.

Strengths

  • Design-time tooling with no runtime dependency on the MCP.
  • Linting catches question patterns Jev handles poorly.
  • Thresholds come from measurement, not guesses.

Caveats

  • Paid API calls are needed for evaluation.
  • Accuracy stays unmeasured until you run evaluation on your data.

Getting started

Build the MCP server and register it, e.g. claude mcp add tenbin -e TYPESAFE_API_KEY=$TYPESAFE_API_KEY -- node .../dist/index.js (or make register-claude).

At a glance

Type
MCP server
Status
beta
Language
TypeScript
Runtime
MCP server + agent skill
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Documentation, an MCP server and a Skill for designing Jev judgments with question linting, batch evaluation and calibration.

    logicrw/awesome-jev-projects

  • MCP server and agent skill for the TypeSafe AI System One API (Jev): decompose…

    wh000wh000/awesome-jev-live

  • MCP server and agent skill for the TypeSafe AI System One API (Jev): decompose a judgment into Choice / Score / Noul questions, lint them, measure on labelled data, and put calibrated thresholds in code.

    hellogumbo/awesome-jev

Frequently asked questions

What does Tenbin produce?

Linted Jev questions, measured thresholds and production code that calls the TypeSafe SDK directly.

Does it need a TypeSafe key?

Not for linting or code generation; only for evaluation runs that call the API.

Does production depend on Tenbin?

No, it is used at design time only.

Related projects

  • jev-pipeline-skill

    Agent skill and CLI that chains Jev question design: find decision calls, check fit, draft, lint, calibrate thresholds and plan a cascade.

    Agent skillprototypePython
  • jev-mcp

    Eval-first MCP server for Jev: prototype questions, map them over many items, and measure accuracy and thresholds on labelled data.

    MCP serverbetaTypeScript
  • jev-mcp

    Rust MCP server for Codex that hands classification, relevance, candidate selection and evaluation to Jev with accept/verify policies.

    MCP serverbetaRust
  • jev-mcp-open-source

    Self-hosted Jev MCP gateway on Cloudflare Workers with nine tools for intent routing, reranking and batch judgments.

    MCP serverbetaTypeScript
  • jev-toolkit

    Small stdio MCP server with ask, rank and verify tools backed by Jev, a local event log and installers for several agent harnesses.

    MCP serverbetaTypeScript
  • hippo-memory

    Local long-term memory for coding agents with an optional Jev reranker that orders recalled memories.

    ProjectproductionTypeScript