Termux-first MCP server for compressed local vector search, context-pack retrieval, KV/cache analysis, and custom Adreno Rusticl/Freedreno/KGSL readiness forensics. P0 Hadamard QJL (Zandieh ICML 2024), P1 2-bit Beta (15.72x compression), P2 ProductQuantizer.
TurboQuant exhibits severe definition quality gaps across most tools. Tool names are inconsistent in verb clarity (e.g., 'turboquant_compress' vs 'turboquant_cost_analyze' mixes action verbs with domain nouns). Descriptions are present but generic and under 100 chars for most tools, providing minimal LLM guidance on when/why to invoke them. Parameter descriptions exist for tools 1 - 7 but are sparse; tools 8 - 12 have no visible schema or parameter documentation. No output schemas are documented. Error handling is absent, no recovery guidance, categorization, or actionable error messages visible in code.
Probe Adreno GPU loader and driver readiness
Probe backend capabilities and readiness
Classify context blocks by volatility and recommend cache tiers
Compress vectors using TurboQuant-inspired compression
Build a compressed context pack from files for local retrieval
Search a compressed context pack for relevant chunks
Analyze Claude Code usage snapshots and compute cache efficiency metrics
5 of 12 tools (turboquant_quantize, turboquant_kv_analyze, turboquant_backend_probe, turboquant_opencl_probe, turboquant_adreno_loader_probe) have no visible input schema or parameter documentation in source code.
Tool descriptions are generic and under 100 characters for most tools (e.g., 'Compress vectors using TurboQuant-inspired compression'). Descriptions lack context on WHEN to use the tool, prerequisites, or how it differs from similar tools. LLMs cannot reliably select between tools with such sparse guidance.
No output schemas are documented for any tool. LLMs cannot plan downstream tool calls or extract required fields without knowing what each tool returns. This violates the pattern:response-shaper requirement.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
Analyze KV cache usage and provide optimization recommendations
Probe OpenCL runtime and device capabilities
Detect cache-busting patterns in context blocks
Quantize vectors using TurboQuant quantization methods
Search for nearest neighbors in compressed vector database
turboquant_vector_search has conflicting/redundant parameter names: 'query_vector' and 'queryVector', 'compressed_database_b64' and 'compressedData', 'top_k' and 'k'. This ambiguity forces LLMs to guess which field to populate, increasing error rates.
No error handling guidance visible in code. Tools lack recovery hints, error categorization (retryable vs fatal), or actionable error messages. An LLM receiving a failure has no guidance on next steps.