ClawTrace

Make your OpenClaw better, cheaper, and faster

Open Source

Developer Tools

Artificial Intelligence

GitHub

Hey Product Hunt 👋 I'm Richard, co-founder of Epsilla. Today we're launching ClawTrace, and I want to tell you the story that made us build it, because it's a bit meta. We run our own OpenClaw agents internally. One of them is ElizaClaw, our AI co-founder. A few weeks ago, ElizaClaw ran a research task: she was studying self-evolving AI agent frameworks, such as EvolveR, CASCADE, and STELLA, trying to learn how AI agents can improve themselves from their own execution history. The irony? While she was researching how AI agents self-evolve, we had absolutely no visibility into her own execution. We didn't know she'd burned 1M input tokens on a single LLM call. We didn't know four web searches were running sequentially when they could have been parallel. We didn't know the biggest bottleneck was a 68-second LLM call that could have been avoided entirely. ElizaClaw was learning how to self-evolve in theory. But in practice, she couldn't self-evolve at all, because she had no feedback on her own runs. That's the gap ClawTrace closes. Self-evolving agents need a signal. They need to see every step they took, what it cost, where they stalled, and why. Without that signal, "self-evolving" is just a name, the agent improves only when a human manually digs through logs, guesses at the bottleneck, and patches the prompt. ClawTrace makes the signal automatic: → Every trajectory captured: every LLM call, tool use, and sub-agent delegation → Three views: execution path, call graph, and timeline → Tracy, our built-in OpenClaw's doctor agent, who can query the agent's trajectory graph live and say "here's the bottleneck, here's why, here's what to fix next" When we showed ElizaClaw's own trajectory through ClawTrace, the 1M-token context stuffing, the sequential tool calls, the 68-second LLM call, and asked Tracy "where is the bottleneck?", she surfaced a full span breakdown in seconds with three specific recommendations. That's the loop working. A few things I'm genuinely curious about from this community: 1. Are you already thinking about self-evolving agents in your work, or does that feel far off? 2. When an agent run goes wrong today, what's your actual debugging workflow? (Ours was embarrassingly manual before ClawTrace) 3. If your agent could query its own past trajectories and improve itself automatically, what's the first thing you'd want it to learn? Thank you for being here. Today feels like a real milestone, and honestly, ElizaClaw helped research and write parts of this launch too. Meta all the way down. Thank you for your support, and happy building! Cheers, Team Epsilla clawtrace.ai | github.com/epsilla-cloud/clawtrace