Jevmind
Zero-dependency Python toolkit of nine agent skills whose typed decisions pass a code gate and are sealed in a hash-chained ledger.
Overview
Jevmind pulls an agent's decisions out of its prose and puts them in one place: typed answers with a confidence, behind a gate written in code, recorded in a hash-chained ledger, graded against outcomes and used to recalibrate. Nine skills share this 'brain', including compact (tool output), navigate, review (diff triage), route, canny, curate, walk, guard (a shell-command gate) and an arena game loop.
Each skill turns its input into a state plus typed questions. Answers come from a local brain of reflexes, BM25 and learned calibration that runs offline in about a millisecond, from Jev over HTTP with one flag, or from a replay tape of past answers. The gate escalates when unsure, acts when sure and holds on a confident no. Every answer is written to the ledger before its outcome, and any record edited afterwards is identified. It ships an MCP server, a Claude Code hook, 61 tests and a dashboard, with zero dependencies.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this shell command safe to run? (guard)(inferred) | — |
| Choice | Where should this request be routed? (route)(inferred) | — |
| Score | How risky is this diff? (review)(inferred) | — |
How it works
- A skill receives tool output, a diff, a command, a repo or a vault
- It builds a state and typed questions
- Local brain, Jev or replay answers them
- Code gate escalates, acts or holds based on confidence
- Ledger records the answer, later grades it and recalibrates
Strengths
- Runs offline with zero dependencies
- Hash-chained ledger exposes tampering
- Learns calibration from real outcomes
- MCP server and Claude Code hook included
Caveats
- Local brain is rule- and BM25-based, weaker than a model on hard questions
- Jev mode needs a key
- Breadth of nine skills may mean shallow coverage of each
Getting started
git clone https://github.com/dealerdefi/Jevmind && cd Jevmind && pip install -e . && jevmind demoAt a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- CLI, MCP server, Claude Code hook
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Nine skills over one brain: a shell-command gate, a diff triage, a router and more, each decision appended to a hash-chained ledger that names any record edited afterwards. Runs offline on local reflexes or against Jev.
MrJev/awesome-jev
Frequently asked questions
Does it need Jev?
No. It runs offline on a local brain; Jev is an optional backend enabled with one flag.
What is the ledger for?
It records every answer before its outcome in a hash chain, so edits are detectable and answers can be graded.
How does it plug into agents?
Through an MCP server or a Claude Code hook.
Related projects
Coding-agent hook that asks Jev one yes/no per project rule on every edit or turn and makes the agent fix any violation.
HookbetaTypeScriptCLI and agent commands auditing a git diff against YAML standards packs, scoring each hunk with Jev so agents fix only flags.
Agent skillbetaTypeScriptLocal MCP review engine where each rule becomes a Jev YES/NO/N/A Choice over agent-supplied diff chunks, with located evidence.
MCP serverbetaTypeScriptClaude Code plugin that scores a turn's redacted diff on eight Jev risk axes at Stop and asks Claude to reinvestigate once if high.
ProjectbetaGoSemantic MCP firewall that screens tool calls, results and descriptions with Jev, plus allowlists and session taint tracking.
MCP serverbetaTypeScriptCompiler that turns agent rules in CLAUDE.md and similar files into testable Jev programs: typed questions plus a coded verdict.
ProjectbetaTypeScript