Skip to main content
Glama

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault
JEV_MODELNoModel or alias. Pin jev-1.13.0 if you have tuned thresholds.jev-latest
JEV_TIMEOUT_MSNoDeadline for one logical call, retries included.30000
JEV_MAX_RETRIESNoRetries after the first attempt, on 429 / 529 / 5xx / network errors.3
TYPESAFE_API_KEYYesBearer token. Missing: the server starts, tools return a clear error.
TYPESAFE_BASE_URLNoAPI base. Point at a proxy or a mock.https://api.typesafe.ai
JEV_AUTO_THRESHOLDNoAt or above this certainty, gate is auto.0.85
JEV_MAX_CONCURRENCYNoParallel requests when jev_rank has to chunk.4
JEV_REVIEW_THRESHOLDNoAt or above this (below auto), gate is review; below it, escalate.0.6

Instructions

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

This server publishes no instructions, or was last inspected before Glama recorded them.

Capabilities

Features and capabilities supported by this server

Protocol revision2025-11-25

CapabilityDetails
tools
{
  "listChanged": true
}
logging
{}

Tools

Functions exposed to the LLM to take actions

NameDescription
jev_evaluateA

Ask Jev — a fast, calibrated judgment model — many typed questions about one shared state; returns probabilities plus a gate computed in code. Use it when no other jev_* tool fits. Writing questions (Jev reads literally):

  • State the exact condition in instructions; put boundary cases in criteria.

  • Criteria accept JSON: Choice options and Noul sides as {what, not_for, examples}; Score levels as {summary, signals}. Put lookalike cases under not_for on the side they would wrongly land on.

  • One judgment per question; split compound ones and combine in code.

  • Batch every question sharing a state into ONE call. They run in parallel and cost only their own tokens, so speculative questions are nearly free.

  • Send only the state the question needs; point at parts by path, e.g. ticket.messages[0].text.

  • Never ask it to count, do arithmetic, or compare dates/numbers — compute those in code and pass the result in. Answers: it selects from your options and never generates text; noul is P(yes), ~0.5 means unsure; choice/score carry confidence; gate is auto/review/escalate. Budget: ~64k tokens state + all questions, ~32k state + longest question.

jev_rankA

Rank files or texts by how well each helps answer a query: a relative Choice over the batch for order, Jev's calibrated yes/no per item for filtering. Pass glob or paths for anything you have not already read — do NOT read files to pass their text. The server reads and chunks them and returns only path:start_line-end_line + score, so a 300-file search costs almost no context either way. Pass candidates (id + text) only for text you already hold: search hits, retrieved passages, tool results. unit: chunk (default) ranks line ranges, file ranks whole files by their best chunk. Prefer file for 'where is X': a header comment can outrank the code it describes. Sensitive files (.env, keys, credentials), binaries and generated output are never read; they come back in skipped. Order comes from a Choice among the strongest candidates: trust the top 1-3 when choice_margin >= 0.10, the tail is by relevance. relevance is P(helps answer the query) and drives min_relevance; low any_relevant means look elsewhere. score_spread (top minus median relevance) below 0.15 means the absolute scores are flat; choice_margin says whether the top pick was clear.

jev_pickA

Choose one candidate from a list you already have — grep hits, symbols, failing test names, file paths — instead of guessing. One Choice over the candidate ids plus a none escape hatch, and one absolute yes/no per candidate. Use it after a search has produced the candidates. It cannot invent a value: pick is always one of your ids or null. probabilities are relative (they sum to 1 across candidates and none); fits is absolute per candidate. Read both: a confident pick whose fits is under 0.3 is the nearest miss, not a match. Up to 240 candidates in one call. Keep each text to what identifies it (the matching line plus a little context); for ranking many long passages use jev_rank.

jev_verifyA

Check up to 100 claims against one body of evidence: supported / contradicted / not_addressed per claim, with a calibrated confidence. Pass evidence_path for anything you have not already read — do NOT read a file in order to pass its text. The server reads it, chunks it if it is large, checks every claim against every chunk and returns line ranges. Use evidence only for text you already hold. Use it before you assert something to the user or write it into a file: hold your draft's claims to the source, or check a summary against the document it summarises. The rubric is strictly literal and closed-world: supported only if the evidence states or entails it. A claim that is true in the world but absent from the evidence comes back not_addressed — which is the answer you want when hunting unsupported assertions. conflicting means one part of the evidence firmly supports it and another firmly contradicts it. Claims should be single, self-contained statements — split compound sentences and resolve pronouns first. all_supported is true only when every claim is supported AND the model was confident about each; treat review/escalate gates as claims a human should look at.

jev_gate_actionA

Advisory pre-flight check on an action you are about to take: judges whether it is destructive, outward-facing or credential-touching, how far it reaches, and how it relates to what the user asked for — then returns allow / confirm / block from a deterministic policy in code. NOT A SECURITY BOUNDARY. It is a judgment layer that catches plausible mistakes, and Jev is not hardened against adversarial text: an action or context written to argue for its own approval can shift the result. Never rely on it to contain untrusted input, and never let allow stand in for a real permission check. Use it just before something you cannot cheaply undo: deleting or overwriting files, git history rewrites, installs, deploys, messages, payments, anything touching an external system. Pass action as the concrete call, not a paraphrase: one line with the tool name and arguments, or the fields {tool, command, file_path, target_paths, …}, which scores scope better. Pass user_request in the user's own words, as a string or {latest, previous}. confirm means ask the user first. block means it looks both unrelated to the request and consequential; re-read the request rather than retrying.

jev_next_stepA

Decide what an agent should do next after a step: continue / retry / change_approach / ask_user / done, with the reasoning signals behind it. Use it when a loop has stalled and you are about to guess: a tool returned an error you are unsure how to read, a search came back thin, you have tried the same thing more than once, or you are about to tell the user you are finished. It is deliberately conservative about done: the verdict is downgraded to continue unless the completion signal comes back a confident yes, so a premature 'task complete' turns into another step instead. retry is capped in code — pass attempts and it becomes change_approach once you have tried enough. Pass result truncated to the part that matters (the error text, the head of the output); a huge dump lowers accuracy. Pass last_step as what you actually ran, and goal as the user's objective rather than the current sub-task. Read reasons before acting: it names every code-level override, which is usually more informative than the verdict itself.

jev_list_modelsA

List the model names and aliases this account can send in the model field, with each one's description and release date. Use it before pinning a version: jev-latest is an alias that moves when a new release ships, so if you have tuned thresholds against one model's calibration, pass the versioned id (e.g. jev-1.13.0) to jev_evaluate's model instead. Versioned ids are accepted whether or not they appear in this list. Costs no tokens.

Prompts

Interactive templates invoked by user choice

NameDescription

No prompts

Resources

Contextual data attached and managed by the client

NameDescription

No resources

TDQS

A4.3/5.0

Scored across 7 tools

Disambiguation4/5

Each tool targets a distinct judgment primitive: evaluate (generic typed questions), rank (ordered relative ordering over a batch), pick (single choice among known candidates), verify (claim support against evidence), gate_action (safety pre-flight), next_step (loop control), and list_models (enumeration). The main soft spot is jev_evaluate, which is an intentionally broad catch-all that could absorb most of the other use cases; the description mitigates this with an explicit 'use only when no other jev_* tool fits' rule, so overlaps are recoverable rather than confusing.

Naming Consistency4/5

All seven names share a jev_ prefix and snake_case, and most begin with a clear verb (evaluate, rank, pick, verify, list_models). The deviations are jev_gate_action and jev_next_step, which read as noun phrases rather than verb_noun, but they remain predictable and readable.

Tool Count5/5

Seven tools is a tight, well-scoped set for a judgment/inference backend, each covering a clearly separate decision shape. Nothing looks redundant or padded.

Completeness4/5

The surface covers the core lifecycle of a judgment layer: ask (evaluate), retrieve and narrow (rank, pick), check claims (verify), guard actions (gate_action), control loops (next_step), and introspect models (list_models). Minor gaps exist — no explicit structured-extraction/classification-over-many-items tool, no cost or budget introspection, and no batch multi-evidence variant of verify — but agents can compose the existing tools to work around these.

Maintenance

ActivityMaintained
ResponsivenessUsually handled