JEV for Claude Code: A Second Opinion on Every Command, Checked

On October 16, 2026 @thegreatest_sv posted a five-minute setup that puts "a model that never writes a single word" in front of every move your Claude Code agent makes: screen each command, check each page the agent reads for prompt injection, scan each commit for secrets, all for "a fraction of a cent per check".

The tool is jev-axi, a command-line client for TypeSafe AI's Jev model. When a post about the Claude Code advisor mentioned JEV last week, we found nothing public to check. Now there is a repository, so we read it, TypeSafe's docs and its privacy policy. The idea is sound and the client is careful. But most checks go to a hosted model, and the project is a few weeks old.

The original post by @thegreatest_sv on X (October 16, 2026)

What it does

Jev is not a chat model. Per the jev-axi README, it "answers typed questions with calibrated probabilities": pick one option, rate on a scale, yes or no. The CLI wires those questions into the hooks that Claude Code and Codex already have:

What the post gets right

• "Routine tests pass for free": yes, routine project commands are decided locally and cost nothing.
• "Every commit scanned for secrets": yes, and that scan stays on your machine.
• "Under half a second" and "a fraction of a cent": the README says about 400 ms and "a few thousandths of a cent" per check (as reported by the project). We found no public price list: typesafe.ai has no pricing page we could open, and the CLI's own default for estimating spend is $0.042 per million input tokens, which it tells you to override "if your plan differs".

What it leaves out

• Your agent's work goes to a third party. Commands that are not routine are sent to TypeSafe's API, along with the contents of any local script the command runs. The supervise hook sends up to 20,000 characters of the diff and 12,000 of output per check. jev-axi redacts credentials and skips files like .env and *.pem, and TypeSafe's privacy policy says it will not train or fine-tune models on your input. For client code under NDA, that may still be a conversation to have first.
• It fails open on errors. If Jev times out or the key is missing, the hook falls back to Claude Code's normal permission flow; only an API rejection blocks. That is the right default for not breaking your work, but it means the safety net has holes when the service is down.
• It is very new. The repository was created on September 16, 2026, has 38 stars, and the npm package had 207 downloads in the week of October 6–12. It is too small for AgentGid's tools catalog for now.

Claude Code already has a version of this

Claude Code's auto mode does the same two jobs inside the product. Anthropic's engineering post describes a server-side prompt-injection probe that scans what the agent reads, and a "transcript classifier" that checks each action before it runs and blocks the ones that go beyond what you asked. There is also a built-in sandbox. So the choice looks like this:

The verdict

• On Claude Code alone, start with auto mode. It covers the command and prompt-injection checks without a second vendor.
• jev-axi adds three things auto mode does not: a local secret scan before every commit, loop and early-stop warnings, and the same safety hook for Codex. The git hook is worth having even if you use nothing else, because it never leaves your machine.
• Treat it as a young project. Read what it sends, set --on-error to ask if you want it to stop rather than fall back, and check the price in your TypeSafe console before you put it on every command.

Sources: the jev-axi README and repository, TypeSafe's documentation and privacy policy, npm and GitHub's API, read on October 17, 2026. We did not run jev-axi ourselves.