Built for the SigNoz hackathon · AI & Agent Observability

Your coding agents
watch themselves.

Claude Code, Codex, OpenCode, Grok and Antigravity each keep their own memory and run blind to each other. Notch gives them one shared brain and one baton — and traces every turn, handoff and memory fold to SigNoz as OpenTelemetry gen_ai spans.

View on GitHub

macOS · Windows · Linux · Android — v0.2.1, free and open source (MIT)

The Notch workspace: one thread with a ship route running across claude-code, codex and opencode
One thread over every agent. Here the ship route is mid-run — claude-code → codex → opencode — with the baton passing between them.

SigNoz is the sink, the source, and the trigger

Most tools ship telemetry and stop there. Notch reads its own spans back out of ClickHouse to score its agents, root-cause its own failures, and act on alerts with no human in the loop.

write

All three signals

Traces as gen_ai.* spans following the OpenTelemetry GenAI conventions, six metric instruments, and structured logs that each carry the trace id of the turn that produced them.

read

Queried back out

A 0–100 health score per agent, a trace waterfall that deep-links into SigNoz, and Triage — where an agent pulls its own spans and root-causes its own failure.

act

Self-healing

A firing SigNoz alert takes the agent out of rotation and moves the baton mid-turn. When it resolves, the baton comes back. SigNoz knows the alert fired — only Notch knows the fleet reacted.

“Why did I fail?” — asked and answered

Codex reading 40 of its own spans back out of SigNoz, finding 25 errors, and explaining exactly which model its auth rejected — and the two healthy turns where it recovered.

Notch Triage: Codex root-causing its own failure from 40 SigNoz spans and 25 errors
Triage · finds the most recent failure and the upstream handoff that led into it, then names the fix.

Eight live views over the running fleet

Every screenshot here is a real capture against a live SigNoz. The numbers are measured, not mocked.

Live fleet — six agents connected to one shared brain
Live fleet · every agent hangs off the one shared brain, with the baton edge marching to whoever holds it.
Metrics dashboard with per-agent tokens, cost and health scores
Metrics · tokens and cost exactly as each agent's own CLI reported them, plus a 0–100 health score.
Self-heal episodes showing quarantine, failover and recovery
Self-heal · one episode end to end: alert fired, agent quarantined, baton moved, held 17s, handed back.
Logs read back out of SigNoz with severity filters and trace ids
Logs · read straight out of ClickHouse. Every line carries the trace of the turn that produced it.
Trace waterfall with a deep link into SigNoz
Replay · scrub to any moment, open the waterfall, jump into SigNoz for the full trace.
Decision explorer showing agent reasoning and alternatives
Decisions · what each agent decided and why, with the alternatives it weighed and rejected.
Metric explorer listing series read back from SigNoz
Metric Explorer · every series Notch exports, queried back with its instrument type, unit and labels.
The shared brain with typed memories attributed to agents
The brain · constraints, failures, decisions and facts, each attributed to the agent that learned it.
5
real agent CLIs driven
3
OTel signals exported
733
tests passing
MIT
open source

Run your fleet where you can see it

Notch is local-first — it drives the CLIs already on your machine and your code never leaves it. Point it at SigNoz and the next turn you run is already traced.