Jevex — Development skill for Claude Code

jevex

One MCP tool that returns the files a coding agent should read.

Ask *where is this actually enforced?* [TypeSafe Jev](https://www.typesafe.wiki/) scores a shortlist. The agent reads those lines instead of grepping the tree. Not a second agent. Not a `jev_evaluate` wrapper.


*Illustration, not a screen recording.*

**agy + Gemini 3.8 Flash · n=16 SWE-bench Verified · 1200s cap:** 160s → **69s**, $8.74 → **$3.13**. Patch file 16/16 both arms.

Fixture packet (toy `session.py`, not the SWE run) — [`examples/session-expiry.json`](examples/session-expiry.json):

{
  "source_of_truth": [{
    "path": "src/auth/session.py",
    "symbol": "assert_session_fresh",
    "range": [32, 35]
  }],
  "status": "sufficient"
}

You still plan and patch. jevex only answers **what to read**.

Numbers

Same cheap agent both arms.

	without jevex	with jevex
Mean time	160s	69s
Agent bill	$8.74	$3.13
Hit the file to patch	16 / 16	16 / 16
Jev	—	~$0.0015 / question

Under a 90s cap the without-arm mostly returned empty JSON (timeout): 1/16 vs 11/16 finished. Separate Claude Code fixture, n=5: 6.8 files → 2.2.

Related: OpenJev, Aider's repo map, Serena, the agent's own grep. jevex is index-then-ask for any MCP agent.

How it works

1. Index whatever a coding agent can already read (py, ts, go, rs, md, …).
2. Ask codebase_investigate.
3. Jev ranks a shortlist. The agent Reads those ranges.

you → coding agent → codebase_investigate → Jev on a shortlist → files + lines → agent reads

Integration

Needs a TypeSafe `JEV_API_KEY`.

git clone https://github.com/jimmyhealer/jevex
cd jevex
pip install -e ".[mcp]"

claude mcp add jevex \
  -e JEV_API_KEY=ts_... \
  -e JEV_EXPLORER_REPO=/absolute/path/to/your/repo \
  -- jevex-mcp

Ask: *where is session expiry enforced?* It shoul