hkcjsyzbcgqq/twitch-viewer-bot-ai

• Changelog / Release History
• Module Catalog
• AI Integration
• Roadmap / Planned Features
• Tips for Best Results
• Performance Benchmarks
• Hotkeys / Controls
• Detection / Bypass Status
• Comparison
• Known Issues
• FAQ

I've been running this thing since the 2023 era when viewer bots were glorified HTTP loops that got nuked in a week. The 2026 build is the first one where the AI layer actually earns its keep instead of being a buzzword bolted onto a socket pool.

This is the whole thing laid out. 43 modules across seven categories. If a module isn't here, it doesn't exist yet — check the roadmap.

The 2026 build is where the AI stops being decoration. Three things changed:

Local inference path. Message generation runs against a small local model by default — no API key, no outbound calls to a third-party LLM, no rate-limit surprises. You can point it at a remote endpoint if you want heavier generation, but the default works offline.

Sentiment-aware pacing. The bot reads the last ~40 messages in chat, scores the mood, and adjusts its own cadence. Hype moments get faster replies; slow lulls get fewer, longer messages. This is the single biggest reason the 2026 build reads less like a script than earlier versions.

Pre-send scoring. Every generated message gets a quality score before it fires. Low-scoring messages (repetitive, off-tone, or contextually wrong) get regenerated or dropped. You'll see the score in the transcript log.

What the AI does not do: it doesn't scrape your viewers, it doesn't post to other channels, and it doesn't touch anything outside the target stream's chat. Keep expectations on the rails.

• Ramp, don't spike. A channel that jumps from 12 to 400 viewers in 90 seconds is a flag. Use the warm-up ramp module and let it build over 20–40 minutes.
• Match your proxies to your audience. Geo-consistency check exists for a reason — a channel with a US audience and 80% EU viewer IPs looks wrong.
• Keep chat quiet at first. Run lurker mode for the first few sessions to establish baseline viewer behavior before you turn on the AI chat layer.
• Use per-account cooldowns, not a global one. A single global rate limit is the fastest way to look mechanical.
• Check the health dashboard before every session. A dead proxy silently drops slots and skews your ramp.
• Export and review your logs weekly. The anomaly self-check flags patterns you'd miss by eye.
• Don't max out the slot count on day one. Start at 30–40% of your target and scale.

Numbers from my own test rig and a couple of community reports. Your mileage depends heavily on proxy quality — cheap proxies will tank everything below.

Honest table. "Bypass" here means the bot's behavior doesn't trip the platform's automated viewer-count and chat heuristics — it is not a guarantee against manual review or future platform changes.

1. Get the build. Head to the project page and grab the latest .exe.
2. Extract. Unzip to a folder you control — avoid Program Files if you want portable mode to work cleanly.
3. Run. Double-click the .exe. First launch walks you through proxy setup and a starter config.

No pip, no npm, no git clone. It's a single executable and a config folder.

That's the 2026 build. The AI layer is the part I'm proudest of — it's the first version where I'd actually run it on my own channel without flinching. If something breaks, the logs are verbose on purpose; paste them in an issue and I'll look. Update when you're notified, keep your proxies clean, and ramp slow.