Every skill available. Few in context. mind-nerve sits between a user request and the host runtime (Claude Code, Codex, Cursor, Gemini CLI, and any MCP-aware client). It reads the request, decides which subset of the available skills / tools / agents / MCP servers is relevant, and hands the host a short top-K list — so the downstream LLM never sees the full library in its system prompt. Library size decouples from token cost.

Current release: v0.3.1 (stable). pip install mind-nerve · github.com/star-ga/mind-nerve

What makes it different

2.6× faster deterministic routing than PyTorch — on the routing/score step (0.58 ms mean / 0.97 ms p95, 12-core CPU bare-metal, native Q16.16 backend), and it returns the byte-identical top-K on every run.
Deterministic by construction — Q16.16 fixed-point + SHA-256 tie-break, byte-identical top-K across runs and across CPU substrate (x86_64 AVX2 / ARM64 NEON), with no IEEE-754 fallback in the inference path. A structural guarantee a float model can't offer.
96.06% top-5 accuracy across 11,922 routing candidates (v1.1-oss).
~99% token reduction on a ~4,400-skill Claude Code catalog — only the top-K load per turn, regardless of catalog size.
Native Q16.16 encoder is the default backend (compiled with mindc 0.10.2); a PyTorch/sentence-transformers path is the transparent fallback on any host without the native library.

Model

Architecture: drop-the-decoder encoder + direct scoring head, sliding- window self-attention (window 256, stride 192), 384-dim embeddings.
Base: distilled from BAAI/bge-small-en-v1.5.
Language: English-only (the encoder vocabulary is an English BPE). UTF-8 input is accepted and validated (byte-level BPE tokenizer), but routing quality is tuned for English intents.
Weights: Phase-1 reference weights, Apache-2.0, auto-downloaded from this repo on first route() call (~150 MB).

Usage

python

from mind_nerve import route
result = route("deploy the staging build", top_k=5)
for r in result.routes:
    print(f"{r.score:.3f}  {r.name}")

Runs on Linux, macOS, and Windows from the same universal wheel. The native Q16.16 encoder ships as a Linux ELF; on macOS/Windows the router transparently falls back to the pure-Python path (native Windows is on the roadmap). GPU is a commercial Pro tier — the open-source release is CPU-only.

Scope & honesty

The 2.6× figure is the routing/score step, not end-to-end route latency (the query-embedding encode step is a shared cost, still being optimized). Every number here is reproducible with the bench in tests/perf/ of the repo (fair equal-thread methodology, published determinism hash).
The evidence chain is tamper-evident (hash-anchored), not signed.
Apache-2.0 source + Phase-1 weights; the wheel additionally bundles a separately licensed native runtime component (see LICENSE.md).