ClawTrace is an open source observability and debugging tool for LLM agents that records every step of an agent run as an inspectable execution tree, letting developers identify where tokens are spent, where errors occur, and where performance degrades.
The Problem
LLM agents that chain tool calls, sub-agents, and model queries can be opaque. When an agent fails, developers have to dig through logs to find which step went wrong, which model call consumed most tokens, or where an infinite retry loop started. Without structured tracing, debugging agentic workflows is slow and error-prone.
How ClawTrace Solves It
ClawTrace records every LLM call, tool invocation, and sub-agent interaction as a structured trace tree. Three views give different lenses on the same run: an execution tree showing parent-child call relationships, a call graph showing how agents and tools connect, and a Gantt timeline for spotting bottlenecks and parallelism. An AI analyst called "Ask Tracy" answers natural language questions about cost and performance. Apache 2.0 licensed; deployed as an OpenClaw plugin by Epsilla.
Key Features
- Execution tree visualization for every LLM call and tool use in a run
- Per-step token usage and cost breakdown with retry inspection
- Gantt chart timeline for identifying bottlenecks and parallel execution
- Call graph view showing relationships between agents, tools, and models
- "Ask Tracy" AI analyst for natural language cost and performance queries
- Apache 2.0 licensed and open source
Who It's For
ClawTrace is best for AI engineers building multi-step LLM agents who need structured observability into token consumption, error chains, and execution timing rather than raw log files.
Compared to LangSmith
Unlike LangSmith, ClawTrace runs as an installable plugin and stores trace data locally rather than sending it to a hosted service. LangSmith offers broader ecosystem support and hosted dashboards; ClawTrace suits teams that want on-premises observability for their agent traces.
