think
bolder
automate
together
JEVMIND.EXE
  • analyze
  • brainstorm
  • set goals
  • take action
  • learn
  • repeat
human
intelligence
large potential
ideas → change → real impact
decide faster
gate in code
seal every answer
grade it later.
AGENT NETWORK
LLMTOOLSDATA
CREATE
PLANDEPLOYSCALE

JevmindTM

human×tools×a brighter tomorrow
watch it work $ git clone https://github.com/dealerdefi/jevbrain read the code ↗
01the idea

One brain,
nine hands.

Jevmind runs on jevbrain: every decision an agent makes becomes a typed answer with a confidence, checked by a gate in code and sealed in a ledger.

write
Text

for a person. Goes to your LLM, or a template.

decide
The brain

pick one · place on a scale · how likely. Typed questions, typed answers, each with a confidence.

act
Code

exact rules, side effects. The gate acts only on an answer it trusts.

61tests passing
9skills
0dependencies
~0.1mslocal decision
2,911decisions in the demo
MITpython 3.10+
02three shapes of answer

Every fork is a typed question.

Three shapes cover every decision. Each comes back with a number that says how sure it is.

Noulhow likely
“This block of output is needed for the task.”
p 0.97 · keep it, byte for byte
Choicepick one
“Who should review this hunk?”
skip
0.00
model
0.00
human
1.00
answer human · confidence 1.00
Scoreplace on a scale
“How risky is this change?”
lowmediumhighcritical
2.96 of 3 · critical

Real answers from the local brain on jevbrain's bundled samples.

03the gate

Code decides. Not the model.

Four rules, written in code. Move the sliders — this is the real logic from mind.py.

the brain answers
its confidence0.80
the gate, set in code0.45
confidence < gateESCALATE
answer says “ask a human”ESCALATE
sure, and the answer says goACT
sure, and it says noHOLD
ACTroute = allow at 0.80
04three brains, one interface

Swap the brain. Keep the gate.

--brain local
default · offline · ~0.1 ms

Small readable rules. Free and the same every time.

--brain jev
TYPESAFE_API_KEY

The same questions, sent to Jev. An unreadable answer gets confidence 0.

--brain replay
a tape

Replays a recorded run to the byte. No network, no bill.

05the nine hands

Nine working tools. One ledger.

Nine tools, one ledger. Open any card and try it — guard and route run live in your browser.

0/9 hands triedtry all nine…
the game

Beat the gate.

Commands fall on the brain. Allow, ask a human or block — before they land. Every answer is checked against the real guard.

score0
combo×1
level1
lives♥♥♥
best0
BEAT THE GATE

A command falls. Decide before it hits the brain.
A / ← allow · S / ↓ ask a human · D / → block

the real guard judges every answer · destroying or leaking is blocked · reaching outside goes to a human
06the ledger

Every answer, before the outcome.

Each line is hashed over the one before it. These six lines are real — check them, then try to cheat.

~/.jevbrain/ledger.jsonl
seqskillsubjectgatehashstatus
07grade, then learn

It keeps a lesson only if it works.

Fit on the older 70% of answers, judged on the newer 30%. No improvement — no lesson.

fit · older 70%
judge · newer 30%

Example: the arena's danger reflex says 0.30, the truth is near 0. learn pulls it down.

$ jevbrain label canny-00007-3fa2b1 supported no
$ jevbrain grade     # Brier score, per skill
$ jevbrain learn     # Platt scaling, kept only if it helps
perfect said 0.30 · true ≈ 0 after learn what it said what was true 0 1
illustration · the arena's danger reflex
08the brain at work

Watch it think.

A visualization of one run: decisions by skill, gate verdicts and confidence — every one a line in the ledger.

JEVMIND TOP · VISUALIZATION
09plug it into your agent

A guard in front of every command.

Every skill as an MCP tool, plus a hook that checks each shell command first.

$ claude mcp add jevbrain -- jevbrain mcp
$ claude mcp add jevbrain -e TYPESAFE_API_KEY=… \
    -- jevbrain mcp --brain jev
{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash",
        "hooks": [{ "type": "command",
                    "command": "jevbrain guard --hook" }] }
    ]
  }
}