vstorm-co/agentcanvas

agentcanvas is a Python library that turns AI agent activity logs into an interactive, visual diagram. When you build a conversational AI agent using Pydantic AI and log its activity to Logfire (a tracing service), agentcanvas reads those logs and generates a single HTML file you can open in any browser to see exactly what the agent did and how much it cost. The output shows the full sequence of events in a block diagram: the user's question, the AI model's decisions, each external tool or function the model called, and the final answer. If your agent uses smaller specialized agents inside it (nested sub-agents), those appear as labeled frames within the diagram, so the structure stays readable no matter how many layers deep the system goes. You can click any step to open an inspector panel showing input and output token counts, exact dollar cost, the model's reasoning, which tools were available, and timing information. For presentations or client meetings, a guided tour mode walks through the diagram step by step with plain-language narration, either automatically or at your own pace using the spacebar or arrow keys. Each conversation turn is displayed in a side panel showing the full back-and-forth transcript between the user and the assistant, including any tool calls in between. Installation is one command: pip install agentcanvas. You provide a Logfire read token, run the command-line tool, and it fetches your most recent agent run and opens a self-contained HTML report in your browser. The file requires no server and works offline, so sharing it is as simple as attaching it to an email. The library can also be imported directly into Python scripts for custom reporting. The project is MIT-licensed and requires Python 3.12 or newer.