$ burnwatch [Live]
Burnwatch
Spend monitoring for AI agents
For developers building AI agents that make autonomous payments
AI agents on x402 and stablecoin rails can make hundreds of payments a minute with no human in the loop. Burnwatch is a smoke detector for those wallets — it mirrors payment metadata (amount, recipient, rail, timestamp), learns each agent's normal behavior, and fires alerts on velocity spikes, drain bursts, unknown payees, new rails, off-hours spend, and more. Observe-only by design: never touches keys, never sits in the payment path.
// features
Velocity spikes, drain bursts, unknown payees, new rails, off-hours spend — 13 transparent detection rules.
Never touches keys or sits in the payment path. If Burnwatch goes down, your agent keeps paying.
Every alert includes human-readable evidence JSON — no black-box ML, no mystery flags.
pip install burnwatch, then one record() call after each payment. Works with x402, stablecoins, and custom rails.
Agents enter a warm-up period before rules fire. The system learns what normal looks like before raising alerts.
Fleet view, per-agent spend charts, alert management, and audit log — all scoped to your org.
// how it works
- 01 Sign up and get a token
Create a free account at burnwatch.dev, then generate an ingest token in Settings.
- 02 Add the SDK
pip install burnwatch, then call bw.record() after each agent payment. Send a test event to verify the pipeline.
- 03 Get alerts
Burnwatch learns your agent's normal spend and fires alerts when something looks wrong — email, Slack, or webhook.
// under the hood
pip install burnwatch, then bw.record() after each payment. Events POST to a multi-tenant ingest endpoint scoped by org token — never in the payment path.
13 transparent detection rules (velocity, drain burst, unknown payee, new rail, off-hours). Agents enter a warm-up window before alerts fire so baselines learn normal first.
Every alert ships human-readable JSON evidence — no black-box scores. Dashboard shows fleet spend, per-agent charts, and audit log scoped to your org.
Burnwatch mirrors metadata only. If ingest goes down, agents keep paying — the SDK fails open on the agent side and queues locally.
// pricing
2 agents, 10k events/mo, all detection rules
20 agents, 100k events/mo, Slack + Discord alerts
Unlimited agents and events, + Microsoft Teams