Power your agent
Perintah milik penerbit sendiri. Bukan endpoint PlugMCP.
Host a copy with our keys and logs.
npx -y [email protected]Claude Code tidak punya tautan instal untuk perintah lokal. Salin perintah di bawah dan jalankan di terminal.
claude mcp add jev-for-all -- npx -y jev-for-all@0.7.1Jalankan di terminal.
{ "mcpServers": { "jev-for-all": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
URL remote atau perintah di mesin Anda? Perbedaan kedua cara instal
Jalankan di PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
BukaDari penerbit
Jev is TypeSafe's System One decision model. It answers a question in roughly 70 to 500 ms with a typed answer and a confidence score. This repo connects Jev to coding agents so they can ask it which skill to load, which tools a step needs, and how to move in the browser.
For the reasoning behind the design, including where Jev wins and where it loses, see When Jev wins.
An OpenCode plugin that routes skill and tool choices through Jev, TypeSafe's System One decision model.
This project is fully experimental. Do not use it in important or production-grade projects.
It will probably not cause harm. It does not edit your code, your git history or your agent's settings (it adds its own config folder and registers itself as a plugin, and that is all). My concern is a different one: token and context integrity. Parts of this plugin pick a model for you, hand work to a cheaper model, or add text to your agent's context. If a guess is wrong, you can spend more tokens than you save, or a cheap model can get a job it handles badly, and you may not notice.
jev-for-all start or switch with /model./loop because it was
not in Jev's skill list. That is fixed, but there are probably more bugs like it.Use a throwaway project first. Run jev-for-all doctor after installing. Leave the hints off. Keep an eye on
your own token use for a few sessions, and compare it with what you normally spend.
Jev runs on OpenRouter's alpha Decisions API. It is not an LLM. You pass it state, it returns typed answers with a calibrated confidence for each. The repo connects it to OpenCode, Claude Code, Hermes and pi (senpi) through one shared decision contract, so every harness gets the same answers.
bunx jev-for-all@latest init # or: npx jev-for-all@latest init
jev-for-all doctor # checks every piece and says how to fix what is wrong
Use @latest: bunx and npx cache the first version they ever downloaded, so a bare bunx jev-for-all init can run an old copy that does not know init. (bun pm cache rm clears it.)
init is safe to re-run. It needs bun (the Claude Code hook, the MCP tool and the
launcher run on it) and does four things:
~/.config/jev-for-all/config.json (backend: your OpenRouter key or a Laya URL, mode 0600)
and ~/.config/jev-for-all/models.json (one entry per tier, see below). Existing files are never
overwritten. The Claude Code plugin also creates them on its first session if they are missing.Pass --key sk-or-... (or enter it when asked) to store the key. An exported OPENROUTER_API_KEY works
without storing anything. Other flags: --laya <url>, --no-claude, --no-opencode, --dry-run.
Do it by hand instead:
/plugin marketplace add emirbartu/jev-for-all # inside Claude Code
/plugin install system-one@jev-for-all
// ~/.config/opencode/opencode.jsonc
{ "plugins": [ { "package": "jev-for-all" } ] }
Let your agent install it: point it at https://raw.githubusercontent.com/emirbartu/jev-for-all/main/docs/install.md.
~/.config/jev-for-all/models.json is created with the defaults, so it doubles as the example:
{
"light": { "harness": "claude", "model": "haiku", "effort": "max" },
"standard": { "harness": "claude", "model": "sonnet", "effort": "auto" },
"heavy": { "harness": "claude", "model": "sonnet", "effort": "auto", "minEffort": "high" }
}
effort is low, medium, high, xhigh or max, or auto to let Jev choose per request.
harness is claude (model alias or id, run with --effort) or opencode (provider/model, run with
#variant). Delete the file to get the defaults back. Plugin options for the OpenCode side are in
Full configuration.
These are the original numbers from a test set the author wrote, so treat them as best case. See Read this first for what real use looked like.
browser_task run, "open this page and click that article", took 2 steps and
3 decisions, cost $0.000199, and finished in 3.7 s on the correct page.gate::mechanical, so they get no skill (skills.mechanicalVeto, default 0.5).browser_task: given a goal, Jev picks each click, target and typed
value.Built for people who split work across cheap subscriptions (a $20 Claude plan plus OpenCode Go).
Three tiers, each a harness, a model and a thinking effort (tierConfig in spec/decisions.json):
jev-for-all start "fix all the eslint errors across the codebase" # light -> claude --model haiku --effort max
jev-for-all start "add input validation to createOrder" # standard -> claude --model sonnet --effort medium
jev-for-all pick "<prompt>" # print the decision and command only
start decides once, from the first prompt, and never switches mid-session (a model change throws away
the prompt cache and the agent's context). One Jev call returns the tier, the effort and whether the
task needs an external service; if it does not and the task is light, a Claude session starts lean
(skills and MCP off). Change any tier in ~/.config/jev-for-all/models.json, for example
{ "heavy": { "model": "opus" }, "light": { "model": "opencode-go/glm-5.3" } }.
The model at the top is whatever you started: Claude Code on Sonnet (or OpenCode). It stays in charge and delegates cheap work to the light tier:
jev-for-all start launches a stronger tier with a jev-light subagent
(--agents, built from models.json, so a config change applies at the next launch), and, with hints on, the hook adds a
one-line "use a subagent with model: haiku" hint when Jev judges a prompt light."light": { "harness": "opencode", "model": "opencode-go/..." }. The plugin then
advertises an MCP tool, jev_delegate_light, that runs opencode run --model <model>#max "<task>" in your
project and returns the report plus git status; it advertises nothing when light is on Claude, so it costs
no tokens then. Needs a working OpenCode Go login (opencode auth login; jev-for-all doctor --deep tests it).model: this is the per-prompt
hint layer, off by default because Claude already delegates well on its own and the hints cost tokens in
our measurements. Turn it on with "hints": true in config.json or SYSTEM_ONE_HINTS=1.jev-for-all opencode-agents prints a jev-light subagent pinned to the
light model when light is on OpenCode (the variant field is unverified against OpenCode's agent schema).bun scripts/bench-claude.ts: 12 small coding tasks, each with a hidden check that a hand-written
reference solution passes, run through headless Claude Code with every plugin off. Costs are the
total_cost_usd Claude Code reports; one run per cell.
Thinking effort barely moves the bill: Sonnet on the same 12 tasks passed 12/12 at both low and high, and high cost 9% more, used 45% more output tokens and took 1.75 times as long, because the roughly 70k tokens of context dominate. Letting Jev choose the effort is a small saving, not a large one.
Read it carefully. These are single-file tasks, so Haiku solving all of them says little about large
multi-file work, and one run per cell is noisy (Sonnet's one failure, lru-ttl, is likely luck).
Tier labels matched mine on 56 of 58 prompts, but I wrote both, and I tuned the light wording on ten of those cases. The default catalog is deliberately
conservative (standard stays on Sonnet); if your tasks look like the benchmark, put Haiku on
standard and keep the 94%. The lean saving is the firmest result: skills and MCP schemas were about
40% of the input for a one-line edit, and it holds on any model.
Laya is an open, self-hosted System One model that speaks the
same POST /v1/systemone protocol as Jev, so it plugs in with one variable and no OpenRouter credits:
uv venv --python 3.12 laya-venv && VIRTUAL_ENV=$PWD/laya-venv uv pip install "laya[serve]"
LAYA_DEVICE=cuda LAYA_PRELOAD=1 LAYA_MODELS=english,multilingual LAYA_DEFAULT_MODEL=english laya-venv/bin/laya-serve
export LAYA_BASE_URL=http://127.0.0.1:8000 # every host in this repo now uses it
export LAYA_MODEL=typed-decisions # optional: english | multilingual | typed-decisions
With LAYA_BASE_URL set Laya is used instead of OpenRouter and never falls back to it, so a stopped
server just means no routing. It is much faster (21 ms median on an RTX 3060, against about 450 ms),
but on our decisions the stock checkpoints are much worse than Jev, measured on the same cases:
Laya's own README says the base checkpoints are meant to be fine-tuned, and that its probabilities are flatter and differently calibrated than Jev's, so thresholds tuned for Jev do not transfer. A wrong skill pick is worse than none, so leave this off until a fine-tuned checkpoint beats the table above. The route to get there is to label a few thousand prompts with Jev once (cents of credits) and fine-tune Laya on them; that has not been done.
browser_task runs one natural-language goal in a real browser. Jev picks every move, and only a
TYPE_TEXT step calls a text model, so a task costs less than a cent. Prerequisites (uv, a
jev-ultrafast checkout, and a Chromium-family browser over CDP) and the MCP wiring are in
adapters/browser-mcp/README.md.
Every model dispatch sends the conversation tail, including tool-result bodies up to
tools.stateBudget (6000) characters, plus the tool catalog's names and descriptions, to
OpenRouter's alpha Decisions endpoint. browser_task also sends the current page's text and
controls on every step while it runs. Nothing is sent without an API key. Every Jev call fails
open: a timeout, error, malformed answer or low-confidence decision changes nothing and never
blocks a model call.
The useful options, with defaults:
apiKey: OpenRouter key, or set OPENROUTER_API_KEYmodel: ~typesafe/jev-latesttimeoutMs: 2500, the per-request timeoutskills.enabled, skills.gateThreshold: true, 0.3 for skill routingtools.enabled, tools.maxTools: true, 12 for tool-subset routingobserve.enabled, observe.file: false, none, per-message usage JSONLbrowser.enabled, browser.jevDir: false, ~/jev-ultrafastcontrol.verify: false, verification-gate hint before a completion claim{
"plugins": [
{
"package": "/path/to/jev-for-all",
"options": {
"apiKey": "sk-or-...", // or set OPENROUTER_API_KEY
"model": "~typesafe/jev-latest",
"timeoutMs": 2500,
"debug": false,
"agents": ["build"],
"serverURL": "https://openrouter.ai",
"skills": { "enabled": true, "rerank": "auto", "gateThreshold": 0.3,
"rerankAbove": 40, "rerankBelowP": 0.5, "shortlist": 3,
"fitsThreshold": 0.3, "minConfidence": 0.3 },
"tools": { "enabled": true, "maxTools": 12, "minToolProbability": 0.05,
"needsToolThreshold": 0.3, "minConfidence": 0.3,
"alwaysVisible": ["read","write","edit","bash","grep","glob"],
"stateBudget": 6000 },
"observe": { "enabled": false, "file": "/tmp/system-one-usage.jsonl", "retain": 20 },
"browser": { "enabled": false, "jevDir": "~/jev-ultrafast",
"maxSteps": 12, "timeoutMs": 180000 },
"control": { "verify": false }
}
}
]
}
bun run typecheck # type check
OPENROUTER_API_KEY=... bun scripts/jev-probe.ts decisions # live Jev smoke test
Jev is a TypeSafe model; this is an independent integration, and the repo is not affiliated with any other organization or team.
An OpenCode plugin that routes skill and tool choices through Jev, TypeSafe's System One decision model.
{ "mcpServers": { "jev-for-all": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}Tool
Belum ada daftar tool. Server remote mendapatkannya setelah pemeriksaan kesehatan pertama.
FAQ
Salin konfigurasi stdio untuk Cursor, Claude Desktop, Claude Code, atau VS Code di halaman ini. Konfigurasi ini memakai instalasi dari penerbit, bukan proxy PlugMCP.
Tidak ada variabel lingkungan wajib yang tercantum di catatan ini.
Jev For All berjalan secara lokal melalui stdio.
Terakhir diperiksa 2026-11-13. An OpenCode plugin that routes skill and tool choices through Jev, TypeSafe's System One decision model.