Compared with other Jev tools

The Jev ecosystem is a few weeks old and moving fast. This page compares what the tools do, from their own READMEs as of 2026-09-27. Stars are a snapshot and mostly show who got listed early. If a tool below fits you better, use it.

At a glance

	jevx	jev-mcp	jev-cli (Nasrallah)	jev-cli (shaharia)	hermes-jev-skills
Form	agent skill + CLI	MCP server	CLI + Claude Code plugin	CLI	skills + Hermes plugin
Stars	new	409	22	30	862
Installs into Claude Code, Codex, ~/.agents	yes, one command	via MCP config	Claude Code plugin	no	Hermes, Claude Code, Codex
Works with no config (key in env)	yes	yes	key in OS keychain	yes	key in Keychain
Any System One endpoint (local, self-hosted)	yes, profiles	TypeSafe / OpenRouter / Cloudflare / Vercel	TypeSafe	TypeSafe	TypeSafe
unsure verdict and its own exit code	yes (exit 3)	act / review bands	--fail-on (exit 2)	abstain band (exit 11)	shadow / on
Error never read as “no”	yes (exit 4)	typed errors	not stated	exit 10 = false, errors separate	not stated
Parallel batch over files / JSONL	yes	batch tools	batch with a pool	batch with back-off, --resume	per skill
Saved questions with their own thresholds	yes (global or per repo)	no	no	YAML question files	no
Context from your AGENTS.md / CLAUDE.md	yes (## Jev section)	no	no	no	memory filtering
Hooks as config (guard, turn check, routing)	yes, plugins, off by default	gate / review tools	compact plugin	no	routing, screening
Reply validation (ids, probabilities)	yes, fails closed	yes, fails closed	not stated	not stated	not stated
Call log without content	yes	not stated	not stated	cost footer per call	dashboard

“Not stated” means the README does not say; it may still exist.

Where each one is the better pick

• You live in MCP clients (Claude Desktop, Cursor, anything MCP): jkudish/jev-mcp. Eleven ready-made tools (verify, screen, rerank, extract, review, gate…), provider fallback across TypeSafe, OpenRouter, Cloudflare and Vercel, and strict reply validation. jevx has no MCP server.
• You want the most task verbs in a CLI: Nasrallah-AL/jev-cli has verify, screen, classify, extract, rerank, route, compact and more, with table / markdown / CSV output and the key in the OS keychain.
• You run large batch jobs: shaharia-lab/jev-cli has --resume, back-off, --dry-run with a cost estimate and YAML question files validated offline. jevx batches restart from zero.
• You use Hermes: kerpopule/hermes-jev-skills is the most complete skill set, with the best key handling (a one-time local page writes it to Keychain) and published notes on what leaves your machine.
• You only want one guard: jev-airbag (PreToolUse for Bash) and jev-belay (Stop, with a published AUROC on 100 labelled stops) do one thing each, with measured results.
• You want to shrink context: fast-jev-compaction (7,003 stars) prunes stale tool calls instead of summarising. jevx does not do compaction; a plugin can exec it.
• You want no hosted model at all: jevless serves the same API from any model with logprobs, and open-jev distils a small local classifier. Point a jevx profile at either.

Where jevx is different

• Agent first. The product is the skill: one install teaches Claude Code, Codex and ~/.agents when to use it, how to word a question, and what to do with each answer, with 18 real scenarios.
• One engine, everything configurable. ask plus four shortcuts; every threshold, retry and timeout in one settings system (built-in < config < profile < question < flag).
• Your context, read only. The ## Jev section of the agent files you already keep, global and per repo.
• Hooks as config. A plugin is a few JSON fields; exec plugs in any tool above without a jevx release.
• Honest by default. unsure is an answer, errors never look like “no”, and plugins start in shadow mode.

What jevx does not have yet

An MCP server, --resume for batches, a cost estimate before a run, keychain storage for keys, an eval command that picks thresholds from your own labelled data, and hook support for agents other than Claude Code.