Claude Code integrations with Jev split into three things: a community mod that decides which Claude model handles each task, several independent MCP servers that give Claude access to Jev’s judgment tools mid-session, and TypeSafe’s own official agent skill. None of the three was built or is maintained by Anthropic.
What the integration does
Each piece does a different job. The community mod, jev-model-router, sits inside Claude Code’s own model routing and picks which model, Haiku, Sonnet, or Opus, handles a subagent or conversation turn. The MCP servers work the other way: Claude calls Jev mid-session and gets a typed judgment back, similar to how the Pi integration exposes a separate model as a tool. TypeSafe’s official skill teaches Claude to write code that calls the Jev model directly, instead of through a tool. Jev works as a classification model here: given text and labels, it returns a probability for each one instead of prose, which is what makes it useful as the decision layer behind a general-purpose AI model like Claude.
Official support
Claude Code’s Jev support splits into three buckets, and only one counts as official. The community mod and every MCP server here are built by outside developers, with no involvement from TypeSafe or Anthropic. TypeSafe’s own skill, typesafe-ai/skills, is the exception: MIT-licensed and shipped directly by TypeSafe, which is the only official Jev support Claude Code has. Anthropic has not built, shipped, or endorsed any Jev integration for Claude Code.
Setup
- To have Claude write Jev-calling code rather than call Jev as a tool, install the official skill:
claude plugin marketplace add typesafe-ai/skills, thenclaude plugin install typesafe@typesafe-ai. For other agents, skills.sh works too:npx skills add typesafe-ai/skills --skill typesafe-ai(-gfor global). Manual guide: docs.typesafe.ai/agent-skill#installation. - Get a
TYPESAFE_API_KEYfrom console.typesafe.ai/settings/keys. See how to get Jev access for account details. - Pick an MCP server from the table and register it.
| Server | Exposes | Install | Key needed |
|---|---|---|---|
| jkudish/jev-mcp | 10 judgment tools: verification, screening, ranking, classify, compare, extract, patch review | claude mcp add jev -- npx -y @jkudish/jev-mcp |
TYPESAFE_API_KEY, or falls back to OpenRouter, Cloudflare Workers AI, or Vercel AI Gateway |
| rashedInt32/jev-mcp | jev_classify, jev_score, jev_check, jev_ask (batched), jev_triage (reads files) |
npx -y jev-mcp, or /plugin marketplace add rashedInt32/jev-mcp then /plugin install jev@jev-mcp |
TYPESAFE_API_KEY or JEV_API_KEY |
| blakestone-x/jev-mcp | classify, score, check, match/screen, each with confidence | claude mcp add --scope user jev -e TYPESAFE_API_KEY=... -- uvx --from git+https://github.com/blakestone-x/jev-mcp@v0.2.1 jev-mcp |
TYPESAFE_API_KEY |
| itsmostafa/typesafe-mcp | Noul, Choice, Score, plus rate limiting and batching | curl -fsSL https://raw.githubusercontent.com/itsmostafa/typesafe-mcp/main/install.sh | sh |
TYPESAFE_API_KEY, or OPENROUTER_API_KEY with Decisions endpoint access |
- Confirm the server registered with
claude mcp list. - jev-model-router is listed on the aitmpl.com catalog under
productivity/jev-model-routerand auto-loads from.claude/skills/jev-model-router/asjev-model-router@skills-dir. Its install command was not independently confirmed, so check davila7/claude-code-templates on GitHub for the current one.
What people built with Jev and Claude Code
The example repos linked below show what developers actually built by wiring Jev into Claude Code, from routing hooks to full MCP servers.
Limits and gotchas
Registering an MCP server with claude mcp add can briefly expose the key as a process-list argument, which blakestone-x/jev-mcp’s README notes, since keeping the key out of tool arguments is the point of that server.
Most servers need a TYPESAFE_API_KEY, but a few degrade instead of failing when it is missing. jev-model-router falls back to the engine’s own $.model.classify, and itsmostafa/typesafe-mcp accepts an OPENROUTER_API_KEY instead, given Decisions endpoint access.
adarshmishra07/jcm-router, a local proxy between Claude Code and the Anthropic API, classifies each message to pick Haiku, Sonnet, Opus, or Fable, plus a reasoning effort level. It routes subagents and parallel tasks freely but leaves the main conversation alone to protect its cached context, and refuses to start without a TYPESAFE_API_KEY in the apikey_... format. Its cost comparison, 309 requests at $106.73 against an $87.19 baseline, is the author’s own measurement, unverified.
Two smaller tools worth a mention: NiazMorshed2007/jev-review for code-quality scoring, and GhalebDweikat/winnow for screening tool output block by block. Neither got the scrutiny here that the main servers and the official skill did, so read a project’s README before handing it an API key.
Routing every subagent spawn through a classifier adds a network round trip, and Jev’s speed and pricing figures set the ceiling on that cost. For when routing to Jev beats asking Claude itself, see Jev versus Claude.
FAQ
Does Claude Code support Jev officially?
Only through TypeSafe’s own agent skill, typesafe-ai/skills, an MIT-licensed skill shipped directly by TypeSafe. Anthropic did not build or endorse this skill, and none of the MCP servers or the community mod on this page carry any official backing from TypeSafe or Anthropic.
Do I need a TypeSafe API key to use Jev with Claude Code?
For most of the MCP servers, yes. jkudish/jev-mcp, blakestone-x/jev-mcp, and itsmostafa/typesafe-mcp need TYPESAFE_API_KEY, from TypeSafe’s console, though itsmostafa’s also accepts OPENROUTER_API_KEY with Decisions endpoint access. rashedInt32/jev-mcp accepts TYPESAFE_API_KEY or JEV_API_KEY. jev-model-router is the exception, falling back to the engine’s own classifier without one.
What does Jev decide inside Claude Code?
It depends on the integration. The MCP servers answer whatever judgment you send, a classification, a score, a comparison, or a pass/fail check. jev-model-router decides which model, Haiku, Sonnet, or Opus, handles a subagent or conversation turn. The official skill has Jev decide nothing itself. It teaches Claude to write code that calls Jev directly.
Is it free to use Jev with Claude Code?
No. Every MCP server and the community mod call TypeSafe’s paid API, which needs a working TYPESAFE_API_KEY and follows TypeSafe’s own usage-based pricing, covered in the speed and pricing guide. The Claude Code side costs nothing, since the skill, the mod, and every MCP server are open source, but calls to Jev itself still cost money at TypeSafe’s rates.
Examples for Claude Code
typesafe-ai/skills (official TypeSafe agent skill)
Official agent skill for Claude Code, Codex and other agent environments; installs via claude plugin marketplace add typesafe-ai/skills or npx skills add typesafe-ai/skills. MIT licensed.
jkudish/jev-mcp
Proof-of-concept MCP server for Jev that lets Claude Code, Claude Desktop and Codex call the model and receive probabilities they can branch on. The most-starred Jev MCP server found.
tamaratran/fast-jev-compaction
Claude Code plugin that replaces the compaction summary with Jev decisions. Every tool call and result is scored in one request against the whole conversation, stale ones are dropped or truncated, and everything kept stays verbatim. Also usable as an npm library.
NiazMorshed2007/jev-review
Local-first MCP plugin giving Claude Code, Codex, Cursor and OpenCode structured code-quality scores (correctness, complexity, modularity, security) while they work. Sends focused diffs to Jev and keeps keys and code local.
jev-router (gargpratyush)
Automatic per-turn model routing for Claude Code and OpenAI Codex. Jev sends simple work to the fast model tier and difficult work to the strong tier while preserving each CLI's native tools, sessions, permissions and authentication.
Dicklesworthstone/skillranker
Rust CLI that hands an agent's live session context to Jev, which compares the available skills and estimates which fit the next step. Ships Claude Code hooks, structured JSON, abstention, a terminal display and local feedback.
GhalebDweikat/winnow
A Claude Code hook that splits large tool results into blocks and asks Jev a yes/no relevance question per block, replacing low-relevance blocks with a short recoverable stub before they enter context.
compozy/yoshi
Local context-pruning proxy for Claude Code and Codex. Jev judges which conversation history is still needed above a configurable size gate, and Yoshi applies the validated omissions before forwarding to Anthropic or OpenAI; the author calls it a proof of concept headed into CompozyOS.
blakestone-x/jev-mcp
Python MCP server exposing Jev as classify, score, check, match and screen tools, with a confidence value on every answer. Ships registration steps for Claude Code, Codex and Cursor that keep the API key out of tool arguments.
lomeshdutta/skill-router
A personal skill router for Claude Code that uses Jev to decide which of your installed skills a session needs, built after the author found they only reached for five of about 90 installed skills.
research_desk
A live market news desk that takes yfinance headlines and returns ranked trade ideas, running 215 typed judgments through Jev per pass in 2.7 seconds for about a quarter of a cent. Built with Claude Code and TypeSafe's agent skill.
AntonioCoppe/jev-harness
TypeScript library that wraps Jev choice, score and noul answers in the parts production needs: a policy, a confidence gate, shadow mode, reusable recipes and an offline eval CLI. On the same row-filter job it measured 1.3 s against 48.9 s for the Claude Code CLI.
adarshmishra07/jcm-router
Local proxy between Claude Code and the Anthropic API that picks the model and effort level per message with a Jev decision. The author measured that routing subagents saves money while routing an established main chat loses it: 309 logged requests cost $106.73 against an $87.19 baseline, and $17.12 of the $19.53 loss came from the main chat alone.
BeLazy167/typesafe-mod
Claude Code mod that routes decisions to Jev: it ranks installed skills against each prompt and answers the agent's own this-or-that questions when confidence is high enough, otherwise falling back to asking the user.
rashedInt32/jev-mcp
MCP server exposing Jev as typed judgment tools (classify, score, check, batched ask) and shipping as a Claude Code plugin. Runs via npx jev-mcp.
leepokai/jev-guard
Security hook for coding agents that risk-scores every tool call with session context (deny/ask/allow), flags prompt injection in tool results and checks skills and plugins. Supports Claude Code, Codex, Copilot, Gemini, Cursor, pi, OpenCode and ACP.
noplan-inc/limpet
Stop hook that keeps a coding agent from finishing early. You write completion rules in plain language, and every time the agent is about to stop, Jev scores the stop against each rule in 0.7 seconds and sends it back to work if a rule is violated. Works with Claude Code and Codex.
TypeSafe Agent Skill (docs)
Official docs for the TypeSafe agent skill, which gives a coding agent context on the three question types, architectural patterns and evaluation practice. Installs via Claude Code plugin, npx skills, or manual copy.