Home

Model Chat CLI — Wiki

╔════════════════════════════════════════════════════════════════════════╗
║   A terminal command center for local AI servers                       ║
║   Discover · Chat · Benchmark · Battle · Probe agentic capability      ║
╚════════════════════════════════════════════════════════════════════════╝

Welcome. This wiki is the deep-dive companion to the README. The README answers what the tool does — the wiki answers how it works, why it was built that way, and how to bend it to your setup.

---

Start here

If you want to…	Go to
Install and run for the first time	Installation · Quick-Start
Understand the network scan	Discovery
Read the streaming / TPS / TTFT internals	Chat
Compare models head-to-head	Arena
Find the best system prompt for your task	Prompt-Arena
Hammer a server with load	Stress-Testing
Score tool-calling capability	Tool-Calling-Benchmark
See how the pieces fit together	Architecture
Look up a function or class	API-Reference
Fix something that broke	Troubleshooting
Add a feature or a new server backend	Contributing

---

Tier-zero facts

• No telemetry. Nothing leaves your machine except the requests you make to your own servers.
• No cloud calls. Discovery, chat, arena, stress tests, and the tool bench all run against models you host.
• OpenAI + Ollama protocols. Anything that speaks /v1/chat/completions or /api/chat works.
• One config file. ~/.model_chat_cache.json — that's it.

---

Five things people miss on first read

1. The chat TPS number is decode-only. The timer starts on the first token, so the reported t/s reflects actual generation speed, not wall-clock with TTFT mixed in. See Chat#performance-metrics.
2. The tool bench has six tiers, not one. Quick (7) is a smoke test. EXTREME (8) includes prompt-injection resistance and conflicting tool sources. See Tool-Calling-Benchmark#difficulty-tiers.
3. Realistic-user mode uses Poisson arrivals, not a fixed RPS. It models bursty real traffic with multi-turn growing context. See Stress-Testing#realistic-user.
4. Arena tournament judging is suite-specific. Coding prompts are scored on correctness, not "creativity" or "tone". See Arena#tournament-judging.
5. Discovery caches healthy servers. First scan is slow, every subsequent launch is instant — re-validation runs in parallel against the cache. See Discovery#caching.

---

License & origin

MIT. Built by @mtecnic for people who actually run local models and want to know what they're doing.