Bronze vs Silver vs Gold
Bouncer works fully offline. Connecting a free AI model makes it smarter about which files a prompt is about and how hard the task is.
Capability matrix
| Capability | 🥉 Level 1 Bronze | 🥈 Level 2 Silver | 🥇 Level 3 Gold |
|---|---|---|---|
| API key needed | none | optional (local) | free key or local GPU |
| Network | none | provider you choose | provider you choose |
| 🔑 Secret & 📜 log blocking | ✅ | ✅ | ✅ |
| 🛑 Opus overkill guard | ✅ regex | ✅ AI | ✅ AI |
| 🧹 Context-bloat handover | ✅ | ✅ | ✅ |
| 🌱 Eco mode (new task → fresh session) | — (never blocks on a guess) | ✅ AI decides | ✅ AI decides |
Git-truth hints (git status, @file) | ✅ | ✅ | ✅ |
| Semantic complexity (TRIVIAL / STANDARD / COMPLEX) | regex | ✅ | ✅ |
Semantic file resolution ("the payment webhook" → src/payment/stripe.ts) | — | ✅ at ≥ 0.90 confidence | ✅ at ≥ 0.80 confidence |
| Rewrite of vague prompts (to clipboard) | — | ✅ | ✅ |
| 🧠 Executive summary in rescues | — | — | ✅ |
Every AI result still passes the 5 safety laws: hallucinated files are stripped, low-confidence guesses are dropped, and any call slower than the time budget (1.5 s by default, 3 s on Cloudflare) falls back to Level 1. The AI is only asked when a decision depends on it, so most prompts never wait for it.
How models are graded
Bouncer grades the model you configure automatically:
| Model | Level |
|---|---|
| (none configured) | 🥉 1 |
Open-weights models under 14B parameters: llama-3.1-8b-instant, qwen2.5-coder:7b, gemma2-9b-it | 🥈 2 |
Open-weights models 14B and up: llama-3.3-70b-versatile, qwen2.5-coder:14b, qwen2.5-coder:32b | 🥇 3 |
Frontier "flash/pro" API models: gemini-2.5-flash, gpt-4o, … | 🥇 3 |
*-mini, *-nano, *-lite variants, and models Bouncer doesn't recognise | 🥈 2 |
Not sure which model? See the measured ⭐ recommended models.
Pick a path (each takes ~30 seconds)
| Best for | Cost | Setup | |
|---|---|---|---|
| Groq · Llama 3.3 70B — recommended #1 | lowest latency | free tier | bouncer config --preset groq-gold --key … |
| Google AI Studio · Gemini 2.5 Flash | big context, strong reasoning | free tier | bouncer config --preset gemini-gold --key … |
| Cloudflare Workers AI · Mistral Small 3.1 | already on Cloudflare | free daily allocation | bouncer config --preset cloudflare-gold --account-id … --key … |
| Ollama · Qwen2.5-Coder | 100% offline, private | your hardware | bouncer config --preset ollama-coder |
Check your level at any time:
bash
bouncer stats