jev-claude

Brings levels 5–9 of disler/ten-levels-of-jev ("upstream") into Claude Code, as hooks and an MCP server. Jev is a small, cheap model that answers yes/no, choice and score questions; these levels use it to route models, gate risky tool calls, decide when to compact, and read files without filling Claude's context.

The upstream code these levels need is copied into vendor/ten-levels/, unchanged and under its MIT license. Its README records the upstream commit it came from.

Data leaves your machine. Every Jev call goes to an external model (OpenRouter by default). The hooks send your prompts, commands, file edits and tool output; the MCP tools send whole file contents. Secret-looking files are refused, but anything else in the project can be sent.

Setup

git clone https://github.com/bravegeek/jev-claude ~/dev/jev-claude && cd ~/dev/jev-claude && npm install

# key (OpenRouter); or export OPENROUTER_API_KEY
mkdir -p ~/.config/jev && read -s k && echo "$k" > ~/.config/jev/openrouter_key && chmod 600 ~/.config/jev/openrouter_key

# levels 5 (subagents), 8-9, one project at a time (run inside the project; private to you, not committed)
cd ~/dev/some-project && claude mcp add --scope local jev -- node ~/dev/jev-claude/mcp/server.ts

# levels 5 (you), 6-7, one project at a time (writes .claude/settings.local.json, git-excluded)
node ~/dev/jev-claude/bin/enable.ts ~/dev/some-project        # --off to remove

Both are opt-in per project on purpose. Avoid --scope user for the MCP server: Claude could then send file contents from any project to OpenRouter. And an "ask" permission rule can't be loosened per project, because ask beats allow wherever the rules live. Opting in per project lets you choose "don't ask again" where you want Jev, and leaves it out of every other project.

Related MCP server: Ultra-Eye

Differences from the pi version

• 5 Hooks cannot change the model, so the suggestion goes to you. It only speaks on the first request or when level 7 says the task changed, since switching mid-task throws away the prompt cache. Jev's questions come from upstream; the thresholds are not. Live, the "least costly model" pick leaned fast and was unsure on hard tasks (powerful at 0.38–0.60), while the effort Score split cleanly (routine ≤ 0.98, hard ≥ 1.36). So in lib/route.ts effort leads: ≥ 1.5 or a confident powerful pick → Opus; < 1.0 with a confident fast pick → Sonnet (Haiku for mechanical subagent work, < 0.75); between → no suggestion, Sonnet for subagents.

• 6C cannot rewrite a built-in tool's output, so the warning is added as context beside it.

• 7 Claude cannot compact itself, and a PreCompact hook cannot change the instructions. So the tier goes to you as a message, and the cut-point pick (7C) is baked into the /compact command it suggests. Lines are raised to 60k/100k/140k tokens (JEV_COMPACT_LINES); upstream's demo lines sit below Claude Code's system prompt.

• 6B exempts Claude Code's own ~/.claude and scratchpad from the outside-the-project check. Their content is still screened.

• 8/9 refuse paths outside the project and likely secret files (.env, *.pem, id_rsa, …). MCP tools skip Claude Code's Read permissions, and the file goes to an external model.

Behaviour

• Fails open. If Jev errors or takes more than 5s (JEV_HOOK_TIMEOUT_MS), the tool call goes ahead. The gate is a signal, not a control; see upstream's level 6 notes.

• Every decision is logged to ~/.cache/jev-claude/decisions.jsonl.

• JEV_GUARD=off, JEV_COMPACT=off, JEV_ROUTE=off, JEV_GATES=A,C switch parts off for a session.

• Each gated tool call adds one Jev call (~300ms).

Tests

npm test runs offline against upstream's mock backend (vendored). It checks the wiring: hook I/O, deny format, fail-open, transcript parsing, path rules, and an MCP round trip. The mock picks by word overlap, so it says nothing about judgment quality.