Dipesh Kumar

AXON turns any AI agent into a glass box. One POST per operation spans, tokens, latency, dollars, failures and your whole fleet becomes a living dashboard: trace waterfalls that build in real time, token burn per model, p95 latency, and a burn meter that flags a bad deploy within 15 minutes. No SDK, no signup, no framework lock-in. Postgres-native and self-hostable. The demo is honest: it streams a clearly-labeled simulated fleet right now, and one curl puts your own agent on it.
Two months ago my agent looped. A flaky search tool retried in a retry of a retry, and by morning it had spent $311 across three models and I found out from the invoice, not from my tooling. Every observability tool I reached for was built for requests, not runs: no parent-child span tree, no cost per model decision, no "which exact step set the money on fire."

So I built the thing I wanted at 2 a.m.: a spine, not a platform. AXON is two Postgres tables and one HTTP endpoint. Agents POST spans; AXON rolls them into runs with tokens, cost and failure state; the dashboard streams it live waterfalls that build while the agent is still thinking, a burn meter that flips red if >15% of the last 15 minutes failed.

The demo you're seeing is honest: the fleet on the page is simulated and labeled as such, because nothing kills trust in dev tooling like fake screenshots. Press "run a simulated agent" and watch a real pipeline execute in real time. Then point your own agent at /api/axon/spans one curl, no account and its runs appear next to the demo fleet under your name, never relabeled.

Roadmap if this lands: an OpenTelemetry adapter, webhook alerts on burn, per-feature budgets, and a "replay" that re-runs failed spans against a fixed prompt. What would make this observability you'd actually self-host? I will be in the comments all day.