Monitor LLMs in production. Feature-level observability for safety teams.
The Enterprise tier that sustains the entire OSS platform. Pipe your production LLM traffic, get real-time dashboards of which features fire, alerts when dangerous ones activate, immutable audit trails for compliance.
Low-latency streaming API. Your LLM calls pass through; features emit on WebSocket. ClickHouse-backed dashboards. Pricing from $2/1M tokens.
Deception, sycophancy, shutdown-resistance, jailbreak-fingerprint features monitored 24/7. Slack / PagerDuty alerts on trigger.
Immutable logs of every feature activation. SOC2, EU AI Act, NIST-ready. "We prove we monitored interpretability continuously."
You pipe production LLM traffic through our SAEs. We emit per-token feature activations to your dashboard. Dangerous features trigger alerts. Everything logged immutably for SOC2 / EU AI Act / NIST AI RMF compliance.
Deploying an LLM today is a black box. "We monitor outputs" is 2024 thinking. Watchtower is the first observability layer that reports why the model said what it said, at feature-level granularity, in real time.
$2 / 1M tokens monitored. Free tier (10M tokens/mo) for researchers. Enterprise tiers with SLAs, custom SAEs, on-prem deploy. Revenue funds the OSS platform indefinitely.
We're taking 3 design partners for the Q4 2026 beta. Priority: safety teams at frontier labs, healthcare AI vendors, and AI-Act-regulated deployments. Tell us your model stack, your compliance target, and the one feature you'd most want monitored.