MCP Tool Shop
Open-source MCP-native developer tools — 65+ public repos, weekly spotlights on tools that actually help people building agent workflows.
This Hugging Face org is home to our applied ML research: fine-tunes, datasets, and evaluation artifacts for models that call tools through the Model Context Protocol instead of answering from prose alone.
Flagship: AI Jam Sessions
ai-jam-sessions teaches an AI to play piano and guitar — and sing — via 46 MCP tools, 120 fully annotated songs, and 6 engines.
- jam-actions-v0 — the training dataset: human MCP tool-use traces plus execution-verified QA over real piano MIDI (public-domain scores via piano-midi.de). 115 rows, CC-BY-SA-3.0.
- jam-ft-v1-qwen25 — five LoRA adapters (Qwen2.5-7B-Instruct) fine-tuned on it for tool-grounded musical QA. Preregistered, confirmatory evaluation: all-seeds mean 0.678 to 0.890 vs. the prompted baseline (p = 0.000039), every seed disclosed, full receipts in the model card.
- jam-actions-eval — the preregistered confirmatory evaluation as an academic project page (the full v0 → v1 → B-1 story, figures, and receipts).
- jam-actions-explorer — an interactive, browser-side demo: play the phrases, read the MCP tool-use traces, and explore the eval charts (zero cost).
- Live demo (ZeroGPU) — run the fine-tune against the prompted baseline on the tool-grounded task, live in your browser.
Everything above is grouped in the AI Jam Sessions collection.
Spotlight
📌 Fine-tuning jam-actions: an honest negative, an underpowered positive, and a powered win — the three-arc story start to finish: why v0 failed, why v1 wasn't enough, and how the frozen artifacts cleared a preregistered bar at proper power.
Why it's built this way
Freeze the eval bar before training. Publish every seed, not best-of-N. Report negative results alongside positive ones — the v0/v1/B-1 provenance chain is documented on the model card and the spotlight.
Links
- Site: mcptoolshop.com
- GitHub: mcp-tool-shop-org
- Collection: AI Jam Sessions