Your AI agent is editing files. This map actually shows which ones.

What it does

It watches your repo for file changes and agent activity, then renders an infinite canvas map showing which components your AI agent claims to be working on versus which files it is actually touching. You get live run cards, a working file tree, and dependency arrows between project components, all served from a single static HTML file and a tiny local Node daemon. It reads the agent’s own PLAN.md to build the map and updates itself as the agent edits that plan or the filesystem.

The interesting bit

The entire stack is one 470-line Node daemon and one static HTML file, yet the real trick is letting the agent maintain its own map: it curates PLAN.md as it works, and agenttrail simply draws the contradictions when the filesystem tells a different story.

Key highlights

• Runs entirely offline: one small Node daemon, one static HTML file, no database, cloud service, or accounts.
• Supports Claude Code, OpenAI Codex, Cursor, or any file-editing agent; Claude gets the richest live run cards via repo-local hooks.
• The map is curated by the agent itself through a simple PLAN.md convention with component IDs, dependency arrows, and task status markers.
• One daemon per repo, with every live board accessible from a single browser tab switcher.
• Handles large repos by capping directory breadth and streaming tiny SSE activity ticks; it tells you when the tree is abridged.

Caveats

• The richest live run cards and plan maintenance are only fully wired for Claude Code; Codex and Cursor get file-watcher activity but no run cards or todo lists.
• The full component map requires opting into a PLAN.md convention and local hooks; without that, you get the live file tree and activity feed, but not the dependency graph.
• On very large repos, the file tree is abridged with per-directory caps.

Verdict

Worth a look if you run Claude Code or Codex locally and want a spatial sense of what the agent is actually doing. Skip it if you are after token-level traces or cost analytics—this is about file changes and plan drift, not LLM telemetry.