jev-use — Claude Code MCP server: fast decisions without text
Hands Jev the steps that need only a choice or a score, not prose. Your agent spends fewer tokens on routine forks.
claude mcp add jev-use -- npx -y jev-useWhat jev-use does
jev-use is an MCP server that routes no-text steps through Jev: yes/no, pick one of several options, rate on a scale. The LLM writes, Jev decides. Questions go out in one batched call, and each answer comes back with a confidence value and an escalate flag.
Inside are two MCP tools, a Jev client and a routing layer that drops some questions before they ever reach Jev. Whatever Jev should not decide returns to the agent with a typed reason. A gate checks shell commands: dangerous ones are denied with a reason, and credentials are stripped from a gated action before it is sent. The plugin form with routing rules and the PreToolUse hook lives in harness/claude-code and harness/codex.
You need a provider key in the environment your agent runs in: TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. Without one the server will not run; JEV_BACKEND=mock gives a keyless dry run. Judged state goes to the provider you configure, and only mock stays local. Run npx -y jev-use doctor to check the wiring.
What it helps with
- pick the next build step without calling the model
- deny a dangerous shell command before it runs
- compact a long transcript down to the messages worth keeping
- rate merge risk before shipping a branch
- keep routine forks out of the agent context
Example prompts for the agent
- Ask Jev whether this branch is safe to merge after a green test run
- Rate the risk of this change from routine to incident
- Gate the command
rm -rf buildand tell me if it should run
How to install jev-use
Platform CLI
claude mcp add jev-use -- npx -y jev-usePlatform CLI
codex mcp add jev-use -- npx -y jev-useConfig file
{
"mcpServers": {
"jev-use": {
"command": "npx",
"args": [
"-y",
"jev-use"
]
}
}
}Config file
{
"servers": {
"jev-use": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"jev-use"
]
}
}
}Platform CLI
gemini mcp add jev-use npx -y jev-useStep by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about jev-use
How do I install jev-use in Claude Code?
Run claude mcp add jev-use -- npx -y jev-use. Then set a provider key in your agent's environment and check the wiring with npx -y jev-use doctor.
What does jev-use do?
It sends Jev the steps that need only a choice or a score rather than text, batching the questions into one call and returning answers with confidence and an escalation flag.
What do I need for jev-use to work?
A key for one provider: TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. For a keyless run, set JEV_BACKEND=mock.
