QVAC-MCP
# QVAC-MCP — shared MCP for Decentralized AI Hackathon (TryDojo / ISD Summit 2026)
Shared Model Context Protocol server used across all 4 track projects + main repo.
Stack: Python + FastMCP + UV. Inference rule: QVAC on-device / P2P only, never cloud.
## Proposed toolset (shared across projects)
| # | Tool | Why (hackathon use) |
|---|------|---------------------|
| 1 | `qvac_docs_fetch(url)` | Crawl/research live QVAC docs (`docs.qvac.tether.io`, `qvac.tether.io`). Every project needs correct on-device API usage for Technical 35%. |
| 2 | `qvac_capability_guide(task)` | Returns minimal working snippet + model hint for a task: text-generation, embeddings, rag, transcription, translation, tts, ocr, vision, image-generation, classification, fine-tuning. Avoids hallucinated APIs. |
| 3 | `qvac_compliance_check(design_notes)` | Static check: flags `openai/chat/completions`, cloud URLs, API keys in inference path. DISQUALIFICATION guard (Art. 10). Pass = on-device/P2P only. |
| 4 | `track_brief(which)` / `track_detail(which)` | Returns parsed brief for `track1`, `track2` (Psy models on edge), `general` (main challenge, alias `main`) + rules digest so every agent plans with right context. |
| 5 | `hackathon_plan(track, idea)` | Returns a 48h plan template weighted by jury criteria (Technical 35 / Innovation 25 / Impact 20 / Design 10 / Completion 10) + deliverables checklist (repo + 5-min ES video). |
| 6 | `judge_feedback(design_notes)` | Self-critique prompt enforcing jury weights + disqualification rules before coding. |
Skills exposed as resources: `skill://planner`, `skill://builder`, `skill://reviewer` — same behaviour in all repos.
## Run
```bash
uv sync
uv run python main.py # stdio (for agents)
uv run python main.py --http # Streamable HTTP on :8000
```
TDQS
Scored across 9 tools
Each tool has a clearly distinct role: qvac_index/fetch/search form a browse-retrieve-filter flow, track_brief/track_detail are explicitly tiered summaries, and the remaining tools (rules, compliance, planning, feedback) target separate workflow stages. Even the two discovery-ish tools (qvac_index and qvac_search) are differentiated by map vs. keyword filter.
All names use snake_case and are readable, but they follow multiple conventions: qvac_* for one group, track_* for another, hackathon_* for another, and a lone judge_feedback. There is no uniform verb_noun pattern, though the prefix grouping helps orient an agent.
Nine tools is well within the ideal 3-15 range and each tool earns its place. The count covers knowledge retrieval, track details, hackathon rules, compliance checking, planning, and self-assessment without redundancy or bloat.
The surface is complete for its apparent purpose: an agent can discover QVAC resources, fetch content, search, get track summaries/details, access rules, verify compliance, generate a plan, and run a final judging self-evaluation. There are no dead ends, and the tool chain supports a full hackathon workflow.