Agent Flow

Agent Flow draws Claude Code and Codex sessions as they run. Agent execution appears as an interactive node graph with tool calls, branching, and return flows, so you can trace the sequence of decisions behind a result, see which steps took the time, and watch how subagents coordinate.

Three entry points share the same viewer: a VS Code extension for VS Code-compatible editors, a standalone web app run from source, and an npx command that opens the visualizer in a browser. Claude Code events arrive through a lightweight local HTTP hook server, which the extension configures the first time you open the panel. Codex sessions are picked up by tailing Codex's session rollout files, which supply tool calls, reasoning, and token counts from Codex's own event stream. Both runtimes are watched by default and shown side by side, narrowed with an editor setting or an environment variable.

Anonymous usage telemetry is opt-out and enabled only in the published npm binary. The extension and the from-source dev server emit nothing. Only aggregate values are sent (session and event counts, duration, OS and architecture, version, distinct model IDs observed, which runtimes were watched, and error class names), and it is turned off with an environment variable.