mind-nerve — Intent-classification preselector for agent runtimes

Open the library, hide the cost. A small, fast classifier sits between a user request and the host runtime (Claude Code, codex, grok, kimi, gemini, MCP hosts — 17 CLI runtimes supported by the installer). It reads the request, decides which subset of available skills/tools/agents/MCPs is relevant, and hands the host a short list — so the downstream LLM never sees the full library in its system prompt.

Result: library size decouples from token cost. A 1,300+ skill hub is reachable for ~2k tokens of announce instead of a ~95k-token bulk listing — and only the top-K bodies are ever loaded per turn.

Current release: v0.3.0b8+ (public beta). PyPI: mind-nerve · Code: github.com/star-ga/mind-nerve.

What's new in the beta line

• Per-prompt routing hook — a UserPromptSubmit hook queries the routing daemon per prompt, projects the relevant skills into the CLI's skills dir (atomic symlink flip), and injects a ranked route table with absolute SKILL.md paths. This is what makes a large hub reachable without being announced.
• mind-nerve acquire — vetted external acquisition: search curated sources (Anthropic's skills repo, the official MCP servers repo, the MCP registry API, GitHub search), fetch into a capped quarantine, run a deterministic fail-closed static security scan (shell-pipe installers, reverse shells, exfiltration collectors, prompt injection incl. MCP tool schemas, archive escapes, credential access, persistence hooks), and install the clean packages into the hub with per-file SHA-256 manifests and a live daemon reindex. Threat model: docs/acquisition.md in the repo.
• Native Q16.16 encoder bundled — the wheel ships libmind_nerve_encoder.so with a real encoder_weights.q16.bin blob; MIND_NERVE_BACKEND=native is the default routing path with a PyTorch fallback.
• Cross-CLI audited — the change set was independently audited by two external CLI agents; all critical/high findings fixed with regression tests.

Usage

pip install mind-nerve

from mind_nerve import route
result = route("git status", top_k=5)
for r in result.routes:
    print(r.score, r.name, r.kind)

The first call auto-downloads this checkpoint into ~/.local/share/mind-nerve/runtime/. To pre-seed or use a custom location, set MIND_NERVE_RUNTIME_DIR=/path/to/your/runtime/.

Wire the per-prompt hook into your CLIs (claude-code, codex, grok, kimi, gemini, +12 more):

mind-nerve-install install --cli all

Acquire a vetted external skill:

mind-nerve acquire search "pdf"
mind-nerve acquire install <url>

Model

• Base model: BAAI/bge-small-en-v1.5 (fine-tuned)
• Loss: MultipleNegativesRankingLoss
• Training: 3 epochs, batch 32, lr 2e-5, max_len 256, seed 1337
• Hardware: single CUDA GPU, 119.5 s wall-clock
• Training date: 2026-05-16T04:00:53Z

Catalog

• Version: v1.1-oss (public-clean — no STARGA-private content in the training corpus)
• Candidate pool: 11,922 routing candidates (skills + tools + agents)
• Corpus hash: 1cd130fa98255241b93aaa2fe6a8086bcbf6fc0627c904008cf48ba9f233536d
• Tokenizer hash: cc2a5502d0fa683c98d59da77af1e4ef3a3812e7e2f345c1d8d7a90bed99d817
• Model hash: 83d4d390469bc1bc6a6cac3b9ab8448dcfcd9ac2ba1ab9fce9348c64012681a6

Metrics (held-out eval, 1,193 pairs)

	Baseline (BGE off-the-shelf)	After Phase 1 fine-tune	Δ
Top-1	0.7527	0.8449	+0.092
Top-5	0.9296	0.9606	+0.031
Top-10	0.9489	0.9707	+0.022

What's in this repo

• checkpoint/ — sentence-transformers checkpoint (model.safetensors + tokenizer + config)
• manifest.json — full provenance (corpus_hash, model_hash, training config, metrics)
• route_table.npy — precomputed catalog embeddings (11,922 × 384, float32)
• route_table.jsonl — catalog metadata (one JSON object per row of route_table.npy)

Status & roadmap

Public beta (v0.3.0b8+). The PyTorch reference path above drives the catalog; the bundled native Q16.16 encoder is the default routing path. The active workstream is the pure-MIND migration: the repo's end state is MIND-only (router core, CLI, daemon, MCP server, hook, installer as a compiled native binary), tracked in ROADMAP.md. mindc is at 0.10.2; the kernel tree compiles under a fail-closed CI gate. Cross-arch bit-identity for CUDA (task #57) remains open — the emit path, not the hardware, is the blocker.

License

This model card and the weights it points at are released under Apache-2.0.

The PyPI wheel mind-nerve bundles a FORTRESS-protected libmindnerve.so whose source remains private (STARGA Commercial). The wheel is Apache-2.0; the bundled binary is the protected runtime layer. The Phase 1 inference path published here does not depend on the protected binary.

For commercial deployments needing per-customer FORTRESS-locked builds of the runtime layer, contact license@star.ga.

Citation

@software{mind_nerve_2026,
  author  = {STARGA, Inc.},
  title   = {mind-nerve: Intent-classification preselector for agent runtimes},
  year    = {2026},
  url     = {https://github.com/star-ga/mind-nerve},
  version = {0.3.0b9},
}