cev
Allows cev to send typed decision questions to OpenAI Codex through the Codex CLI/app-server and use Codex's vote as one of two independent model inputs for answers, confidence, agreement, and escalation decisions.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
cev
A decision tool for coding agents that would rather ask a human than guess.
Agents make small judgment calls all day: merge or review, retry or stop, which bucket, what score. Reasoning each one out in the main session costs context and time, and an agent running unattended has nobody to check its guess.
cev takes the call as a typed question, asks a lean Claude that sees only the facts you pass, and returns one of two things: a confident answer, or escalate: true with what was missing.
0 wrong auto-answers in 300 benchmark cases, run through the same warm path the daemon uses. When the facts weren't there, it escalated.
Same input, same result. A repeated question is replayed from the first answer, in milliseconds.
About 1.7s a decision through a warm local daemon.
About 1.6k input tokens a decision, against 38.9k for the same question asked in a full agent session.
No API key and no dependencies. It runs on the Claude Code login you already have.
Install
In Claude Code:
/plugin install cev --marketplace sshworld/cevThat's all, with nothing to configure. New sessions get:
cev_decide, the decision tool.The command gate, which never gets in your way. For a shell command that is hard to undo (a force push,
rm -r, a deploy or publish) it records what cev thinks in~/.cev/gate.jsonlwhile the command runs as usual. SetCEV_GATE=enforceif you want it to ask you before the ones it isn't sure about.
The daemon starts on the first decision and exits when idle.
Needs Node.js 20+ and a logged-in Claude Code. Codex, plain MCP and the command line are covered in other ways to install.
Related MCP server: jev-mcp
What a decision looks like
The agent sends the facts, the rules and the questions:
{
"state": "PR #142: 3 files, +40/-12. CI green. Touches auth/session.ts. Author is a first-time contributor.",
"rules": ["Anything touching auth needs a maintainer review", "Green CI alone is not enough for first-time contributors"],
"questions": [
{ "id": "action", "type": "choice", "ask": "What should happen to this PR?", "options": ["merge", "review", "close"], "risk": "high" },
{ "id": "cla", "type": "bool", "ask": "Has the author signed the CLA?" }
]
}and gets back this (a real run, 1.8s):
{
"answers": {
"action": { "answer": "review", "conf": 0.93, "escalate": false,
"why": "Touches auth/session.ts. Author is a first-time contributor." },
"cla": { "answer": "abstain", "conf": 0, "escalate": true,
"missing": "STATE does not say whether the author has signed the CLA." }
}
}action is answered, with the quote from state that supports it. cla is not: nothing in state says, so cev hands it back instead of inventing an answer. The agent asks the user, or in an unattended run takes the safe default.
Three question types: choice, bool and score. Details in the reference.
When it escalates
cev auto-answers only when all of these hold. Anything else is escalate: true.
The model did not abstain. Every question has an explicit abstain option, and the model is told never to guess a fact that isn't in
state.The answer has exactly the right shape. A malformed answer counts as no answer.
Confidence clears the bar for the question's
risk: 0.60 forlow, 0.75 formedium, 0.90 forhigh.With the optional second model (
CEV_CODEX=1), Codex agrees.
A failed or timed-out call escalates too. It never falls back to a default answer.
How it runs
Claude Code session ─┐
Claude Code session ─┼─ stdio bridge ─▶ cev daemon (one per machine, starts on demand, exits when idle)
Codex session ───────┘ ├─ warm Claude pool (one fresh process per decision)
├─ queue + concurrency cap
└─ ~/.cev/log.jsonl (every decision, in full)Each decision runs in a fresh Claude process with no tools, no settings, no memory and a short system prompt, so nothing carries over between decisions and nothing from your session leaks in. More in how it runs.
Measured
300 cases with a gold answer, 265 of them built around a trap (rule priority, missing facts, prompt injection in the input, stale or conflicting data). Opus 5.5 alone, 2026-11-10:
path | answered right | answered wrong | escalated |
warm, as the daemon runs it | 241 | 0 | 59 |
cold, one new process per case | 240 | 0 | 60 |
This is what was observed on this benchmark with this model, not a measured error rate for your workload.
The cases are synthetic and were written by Claude models, then checked against Codex.
Before 1.4.0 the warm path got one boundary case wrong in 2 of 4 repeats. The default effort is now
high, where that did not recur in 104 repeats. Details.Sonnet 5.5 answered 1 of the 300 wrong, so Opus stays the default.
Asked three times each, 58 of 60 cases gave the same result every time. None gave two different answers.
Methods, the full tables and what didn't work (a cheaper model first, small local models) are in benchmarks.
When to use it, and when not
Use it for bounded calls with fixed options, especially many in a row or with nobody watching: triage, routing, go/no-go checks, rubric scoring.
Don't use it for open-ended work, or as a safety boundary: the gate only looks at command shapes it knows, and it missed 8 of 20 dangerous commands written after its list. If a rule depends on today's date, put the date in state. In a live conversation, an unsure call is usually better put to the person directly.
Also in the box
Command line:
cev decide request.json, for scripts and scheduled jobs.--require-answerexits 3 when cev escalated, so a script stops before acting;--deadline-msbounds the whole call.Codex as a second opinion:
CEV_CODEX=1. Both models must agree before anything is auto-answered.
All settings, the HTTP endpoints and uninstalling are in the reference. Changes per release are in the changelog.
Limits
Tested on macOS with Node 20 and 26. Linux should work. Windows is untested.
~/.cev/log.jsonlkeeps every request in full, including whatever you put instate.It uses your Claude subscription: each decision is one small Opus call.
Credit
The idea of typed decision questions with calibrated probabilities comes from Jev's System One API by TypeSafe AI. cev is an independent re-creation on local CLIs, not a client for that service.
License
MIT
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Put a decision to a council of top AI models and get one brief: verdict, splits, risks, next steps.
Calibrated judgments for text: yes/no probabilities, picks from your options, or scores.
The decision layer for AI agents: act, escalate or refuse, and every decision comes back signed.
Human-in-the-loop for AI agents. Submit choices, get a human decision.
Related MCP Servers
AlicenseAqualityCmaintenanceEnables multi-model consensus decision-making for high-stakes AI decisions, using independent expert models and a judge to surface disagreements and ground decisions.11121 npm1MIT- AlicenseAqualityCmaintenanceProvides agents with fast, typed, calibrated decision tools for classification, scoring, yes/no checks, and gating risky tool calls.53MIT
- AlicenseNot gradedqualityBmaintenanceProvides coding agents with typed classification, yes/no checks, scoring, ranking, and question-answering tools that return calibrated probabilities for fast, reliable decisions.711 npm62MIT
- AlicenseAqualityCmaintenanceEnables agents to resolve ambiguous decisions by returning calibrated probabilities for each candidate action, along with a gate that signals whether to act on the top option or escalate, all via a cheap local or remote backend.5MIT