What it is
Gateway between an agent and its MCP servers: on each request Jev answers one yes/no per server, then per tool of the kept servers, and only those schemas (plus matching playbooks) are returned to the LLM.
What it does: jev_route → schemas of YES tools; jev_call → proxied tools/call; p-thresholds tuned so a needed server is rarely dropped; skills under skills/<name>/SKILL.md routed the same way.
Why it belongs in the showcase: A concrete MCP semantic-firewall shape with published method and thresholds, not just a README claim.
Caveats: Author says ~1/32 cost and ~201k → ~1k schema tokens vs Sonnet full catalog — author-reported; Sonnet was more accurate in their own table; 2★ at intake.
Evidence
Public MIT README; two-step server → tool routing for large catalogs, single request for small ones; skills / playbooks routed the same way. Author says choosing tools cost ~1/32 of Sonnet with the full catalog, with schema tokens ~201k → ~1k in their benchmark image.
Caveats
- Author says 649 tools / 111 hand-labeled requests, ~1/32 the cost of Sonnet full-catalog, ~201k → ~1k schema tokens — author-reported bench image, not reproduced here
- Author says Sonnet with the full catalog picked every needed tool more often (99% vs 95%) — a recall trade-off
- Thin star signal (2★) at intake
- Hosted Jev sees the request text plus tool names and descriptions
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.