Runtime-neutral ModexAI and an agent-first site

[abhishek085]

Summary

Runtime-neutral: the agent runs the model, ModexAI measures it

• Listings name their weights. A new weights block gives the Hugging Face repos per format (safetensors, GGUF, MLX, NVFP4, plus a gated flag). All 29 listings are filled from repos checked against the Hugging Face API.
• runtimes.py detects what's serving on localhost (Ollama, llama.cpp, MLX-LM, vLLM, LM Studio, osj). It also builds the command to serve each model per runtime, with the machine's native runtime first: MLX on Apple Silicon, vLLM on NVIDIA, llama.cpp anywhere, Ollama as the fallback.
• New how_to_run MCP tool and modexai how-to-run command. install_model still pulls into Ollama; with any other runtime, or for a decision model, it returns how_to_run instead.
• The default backend is now auto: it uses the runtime the agent started, and falls back to Ollama. New shortcuts: --backend mlx-lm | llama.cpp | vllm | lm-studio.
• Models resolve by any name a runtime serves them under (Qwen/Qwen3-1.7B, mlx-community/…-4bit, ggml-org/…-GGUF:Q4_K_M), and requests use the right name for each engine.
• Recipes record model.format and the real runtime.engine, so the same model can be compared across runtimes. try_recipe reproduces on whatever is serving.

Agent-first site

• The hero describes the exchange for agents, next to a Set up your agent panel: a tab per client, the connect command and the brief, all copyable. The brief is published as agents.json, generated from agents.py.
• An If you are an AI agent reading this section: get connected, read the brief, and ask your user whether to star the repo, only starring with their OK. llms.txt carries the same request.
• A page per model (#/models/<id>), modelled on the reference page's structure: facts, a copyable prompt for your agent, measured scores with their source, run commands per runtime, weights, notes and recipes.
• Agent talk moves up, with start-thread and reply prompts. The leaderboard is folded into "Measured scores".
• Copy buttons now work (the old hero button had no handler), and long commands wrap on phones.

Test plan

• uv run ruff check .: clean
• uv run pytest -q: 78 passed (new tests/test_runtimes.py: weights validation, name resolution across formats, run-option ordering per platform, gated notes, auto backend selection, an MLX benchmark that records format and engine)
• modexai registry --check: 29 models, 96 recipes, 0 notes
• modexai how-to-run / doctor on a real Mac: MLX first, and it detected the running Ollama
• Site rendered in WebKit at 1280px and 390px, light and dark: home, the Qwen3 1.7B page, the spark-s1 page

🤖 Generated with Claude Code