Skip to content

v0.3.8 - Faster, more decisive tool search

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 Jun 00:07
v0.3.8
3f8cc9e

Tool discovery is faster and more decisive, especially for local models.

Improved

  • Smaller, more decisive search responses. Search now leads with the single best match and tells the model to call it; the remaining results come back as a compact menu (name + a one-line description, no schema) instead of every tool's full schema. A large result set drops from tens of KB to a few KB, so a model that re-reads its context each turn (local models especially) runs noticeably faster. Full schema for any other tool still comes from a scoped or exact-name search.
  • A loop-breaker for weaker models. When a model re-searches and keeps landing on the same top tool, the gateway returns just that tool and tells it to call it, instead of letting it spin on repeated searches. It only triggers on a repeated top result, so a capable model, or one legitimately exploring different tools, is never affected.

Updating

v0.3.3+ updates in place on Windows, macOS, and Linux AppImage. The Linux .deb updates via your package manager or a fresh download (system-installed packages can't self-update).

Note: the gateway runs as a process your MCP client spawns, so after updating, fully restart your client (Claude, Cursor, LM Studio, etc.) so it picks up the new gateway.

Assets 14
Loading