Live efficiency

Watch an agent send calls with and without Perseus. Point it at your own local model to measure the token savings for real.

Simulated Live · local LLM

📁 Connect folder no folder connected · paste below, or connect the demo/ folder base resolved auto-refresh before each measure

Measure once ↻ Loop Reset idle

Requests go straight from your browser to your endpoint. Enable CORS on the server (e.g. OLLAMA_ORIGINS=*) and prompt tokens are read from each response's usage.prompt_tokens. To generate the resolved context automatically, use the Claude Code setup snippet below.

◐ Set it up with Claude Code · paste this to generate the real baseline / resolved contexts and enable CORS

Set up the Perseus live-efficiency demo in this repo: 1. Install and scaffold: pip install perseus-ctx perseus quickstart # detects the stack, writes .perseus/context.md 2. Produce the two contexts the demo compares, and print both: • BASELINE (unresolved): the raw .perseus/context.md PLUS the full MCP tool list an agent would otherwise carry every call → save to demo/baseline.txt • RESOLVED (Perseus): perseus render .perseus/context.md --output demo/resolved.txt 3. Serve a local OpenAI-compatible model with CORS on so the browser can call it: • Ollama: OLLAMA_ORIGINS='*' ollama serve && ollama pull qwen2.5 endpoint → http://localhost:11434/v1/chat/completions • vLLM: vllm serve <model> --port 8000 --allowed-origins '["*"]' endpoint → http://localhost:8000/v1/chat/completions 4. Keep demo/baseline.txt and demo/resolved.txt in the demo/ folder, and tell me the model name. (Re-run `perseus render` any time to refresh resolved.txt.) Then in Live mode I'll click Connect folder → pick demo/, set the endpoint + model, and hit Measure/Loop. With auto-refresh on, the demo re-reads both files before every measurement · so regenerating them here flows straight through, reading the real usage.prompt_tokens from my own model.