Wiring Jev into Claude Code and Codex via MCP

The most common way people first meet Jev is not by writing an API client. It is by hanging it off a coding agent that was already making too many expensive judgments.

Updated September 2026 · 6 min read

Why MCP is the right shape

MCP (Model Context Protocol) is how agents discover external tools. The reason it fits Jev so well is that a decision model is a tool: you hand it a state and a question, you get a typed value back. There is no conversational wrapper needed and nothing to summarize.

Practically, the jev-mcp server exposes Jev as a callable tool, so an agent that already speaks MCP can ask a judgment question without you writing any integration code.

Where it slots into a coding session

A coding agent already makes the three judgments TypeSafe describes, whether or not you have instrumented them:

Moment in the session	The question	Without Jev
Route to the right model	Grade a request's complexity and send it to a fast, balanced, or frontier model.	A frontier call, a hardcoded rule, or nothing at all
Guardrail tool calls	Before an agent runs a shell command or edits a file, score the risk and allow, confirm, or block.	A frontier call, a hardcoded rule, or nothing at all
Compact context	Decide keep / truncate / drop for each stale tool result so long sessions shrink without a lossy rewrite.	A frontier call, a hardcoded rule, or nothing at all

The last column is the honest one. Most agents do not have a guardrail on shell commands because a frontier call per command is absurd and a regex list is unmaintainable. The decision model is what makes the third option — actually checking — affordable.

What is worth gating

Destructive commands. Ask a noul before running anything that writes outside the working tree, force-pushes, or deletes. Branch to allow / confirm / block.
Context triage. Long sessions accumulate stale tool output. A per-item choice of keep / truncate / drop shrinks the window without a lossy summary pass.
Model routing. Grade the task before choosing an executor, so trivial edits stop paying frontier prices.
Do not gate the writing. Code, explanations and commit messages are text. That is the agent's job — see the division of labour.

Fail closed

If the guardrail call fails — no key, no network, timeout — the safe default is to ask the human, not to proceed. A guardrail that silently disables itself is worse than no guardrail, because you will trust it.

Coding and open-source builds here

Listings Jev filed under Coding & Code Review or Open Source:

No coding or open-source listings yet.

What we have not verified

Read the vendor docs, not a blog post

We have not independently verified the current install steps for any specific agent, and they are moving fast — several walkthroughs in circulation differ on where credentials are read from and which config file the hook goes in. Treat any step-by-step you find, including ours, as a starting point and confirm against TypeSafe's own documentation and the MCP server's repository before relying on it.

What we can say with confidence is the shape: a credentials environment variable the agent process can read, an MCP server entry the agent can call, and — for gating — a hook that runs before a tool executes. Anything more specific than that is worth checking at the source.

Common questions

What is jev-mcp?

An MCP server that exposes Jev as a callable tool, so an agent that speaks MCP can ask Jev for a typed decision without custom integration code.

Does Jev work with Claude Code and Codex?

Yes, through MCP plus a hook for gating tool calls. The ecosystem is moving quickly, so verify the current setup steps against the vendor's documentation and the MCP server's repository.

What should I gate with Jev in a coding agent?

Destructive shell commands and context triage are the highest-value gates. Leave code generation, explanations and commit messages to the agent — those are text tasks Jev cannot do.

What happens if the guardrail call fails?

Design it to fail closed. If the decision cannot be made, ask the human rather than proceeding — a guardrail that silently disables itself is worse than none.

BuiltOnJev is an independent community project, not affiliated with TypeSafe AI. Specs and eval numbers come from TypeSafe's published materials; directory figures come from this site's own submissions and are updated as builds arrive.