Jev MCP Server: Fast Fact-Checks and Rankings for Your Agent

• Checks claims against the evidence you give

  Takes a list of claims plus source text and marks each one verified, contradicted or unsupported. It also gives a confidence and says whether the verdict can stand or needs a person to look.

  Try: “Verify each claim in this report against the cited sources”

  Read · jev_verify

• Checks text for hidden instructions before you read it

  Scores fetched or pasted text for prompt injection aimed at an AI, for real content, and for fit with your purpose. It then suggests pass, review, block or skip, and leaves the choice to the assistant.

  Try: “Screen this page for injected instructions before you summarise it”

  Read · jev_screen

• Gives a probability for a statement

  Returns how likely each stated proposition is, from one batched call. Context you add informs the number but does not prove it.

  Try: “How likely is it that this migration is safe to run twice?”

  Read · jev_noul

• Picks the best match from a list

  Scores every candidate against a plain-language question and also checks whether any of them answers it at all, so a weak top hit is not mistaken for an answer.

  Try: “Which of these notes covers rate limits?”

  Read · jev_find

• Scores and sorts every candidate

  Gives each candidate its own relevance score for a query and returns them in order, so the whole ranking is kept and not only the top hit.

  Try: “Rerank these search results for "refund policy"”

  Read · jev_rerank

• Sorts items into your labels

  Assigns each item to one class from a catalog you provide. It flags the ones it is unsure about for review, using a confidence and a gap to the runner-up.

  Try: “Label these support emails as billing, bug or other”

  Read · jev_classify

• Settles a choice between 2 to 6 options

  Weighs the options against your evidence and priorities and returns a spread of probabilities. It can also answer ask the user or investigate when it cannot decide.

  Try: “Which of these three caching designs fits our priorities?”

  Read · jev_decide

• Judges how two passages relate

  Says whether two passages state the same fact, contradict each other or cover different facts, optionally per aspect such as price or date. Agreement between them is not proof that either is true.

  Try: “Do these two release notes disagree about the price?”

  Read · jev_compare

• Pulls values out of a document

  Your patterns find candidate strings and Jev picks the right one for each field. The result is returned word for word from the text, and ambiguous picks are flagged.

  Try: “Extract the invoice date and total from this document”

  Read · jev_extract

• Checks extracted values against their source

  Reviews values you already extracted for invented, off-target, incomplete or wrongly formatted entries, and raises the whole audit if any value looks wrong. It works on text only.

  Try: “Audit these extracted fields against the source text”

  Read · jev_audit

• Scores a proposed code change

  Rates a diff for correctness, match to the request, missing tests and blast radius, and suggests auto, review or escalate. It does not apply the patch or run tests.

  Try: “Review this diff against the ticket before we merge”

  Read · jev_review

• Reviews a patch and its completion claims

  Reviews a patch and checks claims like "tests pass" against the evidence you supply, in one call. It passes automatically only when both hold.

  Try: “Gate this merge: check the patch and the claim that tests pass”

  Read · jev_gate

Watch out for

• Your text is sent to a model provider

  Claims, evidence, diffs and pages you pass to a tool go to TypeSafe or the provider you configured. Do not send secrets or private code you cannot share with that provider. github.com

• Verdicts are probabilities, not proof

  The tool descriptions say a high probability is not proof and that a same-fact verdict only means two passages agree with each other. Treat review results as a second opinion. unpkg.com

• The optional web mode listens on a port

  With --http the server binds to 127.0.0.1 by default and needs an auth token before it will bind to another address. Keep it on loopback unless you add TLS and a token. unpkg.com

Security advisories