Zero instrumentation
Hook-based capture. No SDK, no proxy, no MITM layer.
v0.4 MIT Claude Code
See every agent, tool call, and token as it happens.
Replay sessions deterministically. Zero instrumentation.
Hook-based capture. No SDK, no proxy, no MITM layer.
Pure reducer over append-only log. Deterministic at any timestamp.
Understands subagent types, team markers, depth levels, and iteration counts.
Single binary, writes to /tmp. Your data never leaves your machine.
Live directed-acyclic-graph with depth-grid layout, updated as events arrive.
No SDK. No proxy. No cloud. No setup.
Feature-by-feature breakdown across popular observability tools for AI agents.
| Tool | Zero-instrument | Hook-based | Time-travel | Claude native | Local-only | Real-time graph |
|---|---|---|---|---|---|---|
| Tarsa winner | ||||||
| langfuse | ||||||
| agentops | ||||||
| ccusage | ||||||
| claude-trace | ||||||
| OTel |
Everything you need to understand a multi-agent Claude Code session.
Directed-acyclic-graph of all agents with depth-grid layout. Color-coded edges, updated live as events stream in.
Parallel-bar grouping with depth indentation shows concurrent agents. Mini-map overview for long sessions.
Drag through history to reconstruct exact agent state at any moment. Deterministic via pure reducer.
Per-agent token and cost breakdown. Trace spend back to the exact subagent that incurred it.
Agent-type baselines and z-score analysis surface stalled agents. Root-cause hints included.
Side-by-side comparison of two sessions. Identify regressions, improvements, or behavioral drift.
curl -fsSL https://raw.githubusercontent.com/elcronos/tarsa/main/install.sh | sh
macOS · Linux · WSL2 · ~30s install
A straight pipeline from Claude Code hooks to your browser — no intermediaries.