Image: GitHub
Why it mattersAnyone running several coding agents at once currently has no single view of what they are doing, and this gives one without sending repository activity to a third party.
agenttrail is a local tool that shows what your AI coding agents are doing, across every repository, on one zoomable map. It gained 632 stars on GitHub over the past 30 days. The licence is MIT.
What it does
Each repository becomes a region on a single live map. Inside it, the tool shows plans, tool calls, file changes, and progress as they happen. The project describes support for Claude Code, OpenAI Codex, Cursor, and in its words "any agent that edits files."
The depth of support differs by agent. The project's own table lists live activity through a file watcher for all three, run cards through local hooks for Claude Code only, and map maintenance through CLAUDE.md for Claude Code and AGENTS.md for Codex and Cursor. So the file-change view works everywhere, and the richer per-run detail is currently a Claude Code feature.
How it is built
The design choice worth noting is how little there is of it. The project describes the daemon as one dependency-free Node file of about 470 lines, with a static HTML interface in front of it. It binds only to 127.0.0.1, stores nothing in an external service, requires no account, and sends no telemetry.
It also never writes code. The project is explicit that the tool only observes file changes and agent activity. On size, the project says it has been tested on repositories of 78,000 files, using breadth-first tree reading with a cap per directory and rate limiting on activity.
The problem this addresses is a recent one. Running one coding agent is easy to follow, because you are watching it. Running four across different repositories means the work is happening in terminals you are not looking at, and the first sign that one made the wrong changes is often a diff you did not expect. A shared view of what each session is currently changing is the cheapest way to notice that early.
The local-only design means a team can adopt it without first asking for approval. A tool that watches every repository on a developer's machine and sends that activity somewhere needs a security review. One that binds to localhost, keeps no account, and has no dependencies is something an engineer can try this afternoon and remove completely.
Judge the interest honestly. 632 stars in 30 days is real interest and a real number, and it is still an early project: 102 commits, no tagged releases, and no independent report of it running inside a team for a sustained period. The cost of finding out is low, which is the main argument for looking now rather than waiting.
Source
Primary source: agenttrail on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.