What it is
Jev MCP, by Joey Kudish, turns TypeSafe's Jev decision model into twelve MCP tools that return probabilities rather than prose. They cover claim verification against evidence (jev_verify), screening fetched text for prompt injection before the agent reads it (jev_screen), picking or reranking candidates by meaning (jev_find, jev_rerank), batch classification, bounded choices, passage comparison, field extraction and auditing, and scoring a diff or gating a merge on its completion claims (jev_review, jev_gate).
It runs with npx -y @jkudish/jev-mcp over stdio, or as a stateless HTTP server for a team, which requires a bearer token unless it is bound to localhost. It needs Node.js 22 and a Jev key from TypeSafe, OpenRouter, Cloudflare Workers AI or Vercel AI Gateway, or any endpoint that speaks the same System One API. An agent skill in the package tells coding agents when to call each tool.
The repository has about 550 GitHub stars and an MIT licence; version 0.14.5 shipped on 21 October 2026. A draft pull request opened on 7 October would rename the project to discern-mcp 1.0, list discern_* tool names by default and keep the jev_* names as aliases; it is waiting on an unpublished SDK.
Install
npx -y @jkudish/jev-mcpOur review
Most Jev wrappers expose the model's three question types and leave the agent to compose them. Jev MCP ships finished checks instead, and that is its real value. jev_screen returns an injection probability for a web page before it enters the context window; jev_verify checks each claim in a report against cited evidence; jev_gate reviews a patch and its 'tests pass' claims in one call. TypeSafe bills only input tokens, at $0.042 per million, so the README's own jev_review example (855 input tokens) costs well under a hundredth of a cent.
The engineering is careful. Answers that are missing or malformed fail closed per item and are flagged for review rather than passed through. Retries are limited to 408, 409, 429 and 5xx responses, and connection resets are never retried, because a re-send without an idempotency key could bill a call twice. HTTP mode refuses to serve beyond localhost without a token, caps request bodies at 4 MiB and sheds load past 16 concurrent requests, though it speaks plain HTTP and leaves TLS to your proxy.
The weak point is Jev itself. It is a closed, hosted model, so every page, diff or claim you screen goes to TypeSafe or another carrier. TypeSafe's privacy policy says it will not train on inputs but keeps data 'as long as reasonably necessary', and we found no SOC 2 report. More awkward for jev_screen, TypeSafe's own notes on Jev 1.13 list adversarial content as a known failure mode: injected instructions can move the answer. A screen is a filter, not a guarantee, and the README is honest that its composed tools are not independently benchmarked.
The project also moves fast. It went from 0.1 to 0.14 in about a month, the author calls it early software, and a pending rename to discern-mcp will change package and tool names, with aliases promised for the old ones. Twelve tool definitions are a fixed cost in every session, which is where the single-tool System One Connector is lighter, and Laya's server is the only route that keeps the text on your own hardware.
Where it earns its place
A solo developer running Claude Code with web fetch enabled
Stop injected instructions in fetched docs pages from reaching the agent
Call jev_screen on each page before it enters context and honour its block (0.75) and review (0.25) thresholds. At $0.042 per million input tokens, screening a thousand 5,000-token pages costs about 21 cents, though TypeSafe warns adversarial text can still move the score.
A platform team of eight running agents that open pull requests
Catch 'all tests pass' claims that the logs do not support before a human reviews
Run the HTTP server behind a TLS proxy with a bearer token and have each agent call jev_gate on its patch, test log and claims. Contradicted or unsupported claims come back as reason codes, so reviewers only open the escalated ones.
An operations lead at a 30-person SaaS company
Check that the help centre, changelog and pricing page agree after each release
jev_compare judges paired passages per aspect (price, date, limit) and flags contradictions; jev_extract pulls the values verbatim. The bill is input tokens only, but every page is sent to TypeSafe or the chosen carrier.
Before you commit
- Setup
- claude mcp add jev -- npx -y @jkudish/jev-mcp with TYPESAFE_API_KEY (or an OpenRouter, Cloudflare or Vercel key) set; Node.js 22 or newer. Documented for Claude Code, Codex, OpenCode and Amp, and a first call takes a few minutes once you have a key.
- Integrations
- Twelve tools: jev_verify, jev_screen, jev_noul, jev_find, jev_rerank, jev_classify, jev_decide, jev_compare, jev_extract, jev_audit, jev_review and jev_gate. All are judgments that read inputs and return probabilities; none writes to your systems. A PreToolUse hook example gates agent tool calls against a written policy.
- Data and privacy
- Every input is sent to the Jev carrier you configure. TypeSafe's privacy policy says it will not train on inputs and retains data as long as reasonably necessary; JEV_VERCEL_ZERO_DATA_RETENTION routes Vercel calls only through zero-retention providers on Vercel Pro and Enterprise. HTTP mode requires a token beyond loopback, and the tool screens untrusted content rather than fetching it.
- Cost at scale
- TypeSafe lists $42 per billion input tokens with output free, so 50 million tokens of screened pages a day costs about $2.10. TypeSafe publishes no numeric rate limits, only 429 and 529 responses; the server retries up to 3 attempts by default.
- Limits
- Jev 1.14 takes 64k tokens per request and 32k per question by default, or 262k with a long-context header. TypeSafe's notes say it reads literally, cannot count or do arithmetic reliably and can be steered by adversarial text; the server's tool names will change with the discern-mcp rename.
What's good
- Twelve ready-made checks, including prompt-injection screening and merge gates
- Fails closed on malformed answers; never retries an ambiguous paid call
- Works with TypeSafe, OpenRouter, Cloudflare, Vercel or any compatible endpoint
- Stdio, or authenticated HTTP with concurrency and body-size caps
- Input-only billing makes per-page screening cheap
What's not
- Every input goes to a closed hosted model; no SOC 2 report found
- TypeSafe says adversarial text can steer Jev, which limits jev_screen
- Composite tools' accuracy is not independently measured
- Pending rename to discern-mcp; behaviour still changing
Best for
Developers who want a coding agent to screen fetched content, verify claims and review its own diffs cheaply before calling a task done.
Alternatives
Jev MCP: common questions
- Is Jev MCP worth it?
- Twelve finished checks for the work coding agents skip, at a price measured in thousandths of a cent, built by someone who clearly expects the model to be wrong sometimes. We rate it 4.3 out of 5.
- How much does Jev MCP cost?
- The server is free (MIT). Jev calls are billed by your provider: TypeSafe charges $42 per billion input tokens ($0.042 per million) with output tokens free, and the README says OpenRouter charges the same.
- What are the downsides of Jev MCP?
- Every input goes to a closed hosted model; no SOC 2 report found; typeSafe says adversarial text can steer Jev, which limits jev_screen; composite tools' accuracy is not independently measured; pending rename to discern-mcp; behaviour still changing.
- Who is Jev MCP best for?
- Developers who want a coding agent to screen fetched content, verify claims and review its own diffs cheaply before calling a task done.
- How much does Jev MCP cost for a team or in production?
- TypeSafe lists $42 per billion input tokens with output free, so 50 million tokens of screened pages a day costs about $2.10. TypeSafe publishes no numeric rate limits, only 429 and 529 responses; the server retries up to 3 attempts by default.
- Is Jev MCP safe for company data?
- Every input is sent to the Jev carrier you configure. TypeSafe's privacy policy says it will not train on inputs and retains data as long as reasonably necessary; JEV_VERCEL_ZERO_DATA_RETENTION routes Vercel calls only through zero-retention providers on Vercel Pro and Enterprise. HTTP mode requires a token beyond loopback, and the tool screens untrusted content rather than fetching it.
- What are the best Jev MCP alternatives?
- The closest alternatives we have reviewed are System One Connector, Laya MCP Server, Jev on Viorant and Jev connector for AnythingMCP.
Answers checked Oct 23, 2026.
Make Jev MCP? Show this rating on your site
