Our take
Newless than six months old and changing fast- Activity
- 69
- Community
- 4
- Issues
- 21
- Pull requests
- 62
+224 stars in the last 15 days
- Fewer contributors than 91% of the projects we track
Agenttrail is a focused, local companion for observing coding-agent work, but its young, rapidly changing project and single-contributor base make it better suited to experimentation than as a mature operational dependency.
Good fit if
- You want a local visual overview of file changes and planned project components, or a 3D view of available agent tasks.
- Your work uses local Codex or Claude Code sessions and you are comfortable with an experimental tool whose provider coverage differs between Map and Kitchen.
- You want a MIT-licensed tool that does not require a cloud account or agent-control layer.
Look elsewhere if
- You need a stable, broadly validated product: the project is new, most development depends on one contributor, and Windows and native Cursor behavior remain unverified.
- You need complete remote-session visibility, full model traces, token billing, or an editor extension.
- You need Map to read newer Claude TaskCreate and TaskUpdate lists, or need handoffs and shared artifact receipts without extra metadata.
All health signals
| Last commit | 2026-09-09 (1 month ago) |
|---|---|
| Commits, last 90 days | 100+ |
| Releases, last 12 months | 3 (latest kitchen-v0.1.0-alpha.3, 2026-09-09, pre-release) |
| Contributors | 1 |
| Issues closed, last 90 days | 0 |
| Pull requests merged, last 90 days | 5 (typically merged in 0 days) |
| Project age | 58 days |
Checked on 2026-10-19 with the GitHub API.
Overview
Agenttrail is a local observability tool for AI coding agents. It turns available evidence from local files, plans, and supported agent activity into two browser views: Map, for project structure and changing components, and Kitchen, a 3D task and role view built with Three.js. The goal is to make agent work easier to follow without juggling terminal windows.
It is aimed at developers using tools such as Codex and Claude Code, with optional Cursor hooks. Agenttrail observes rather than launches, directs, or approves agent work. Map and Kitchen are separate local services with different activity coverage; basic use requires no account or cloud service.
Key Features
- Map connects a
PLAN.mdproject outline with observed file changes, dependencies, and session trails. - Kitchen visualizes available native todos, role contributions, and reported completion in a 3D scene.
- Local Codex and Claude activity can be discovered from available logs; optional hooks add supported events.
- File activity can still appear when an agent has no supported native session adapter.
- A scripted example lets users explore Kitchen without an active coding session.
- Both views run locally and do not send prompts, control agents, or require an API key.
Use Cases
- A developer working across several components can use Map to see which planned areas are changing and where dependencies connect.
- A team using Codex or Claude Code can use Kitchen to inspect available tasks and contributions without treating each visual role as a separate agent.
- A developer evaluating an existing repository can run Kitchen to see whether local agent activity is available, without first adding a plan.
- A project using unsupported or remote agents can still use file observation, while recognizing that native tasks and sessions may not appear.
What you need
Detected in the repository
- Node.js >=20 (from package.json)
- Automated checks on GitHub Actions
License in plain words
MITpermissive
- Commercial use: yes
- Modify and redistribute: yes
- You must keep: the copyright and license notice
- Share your changes: no
A summary, not legal advice: the LICENSE file is what applies.
Getting Started
Kitchen requires Node.js 20 or later. From the repository you want to watch, run:
npx agenttrail-kitchen .
For the project map, run npx agenttrail. See the README for setup options, supported integrations, and limitations.
Alternatives
- agent-observability: It collects OpenTelemetry from Claude Code and Codex and visualizes it in Grafana, rather than mapping local project files and activity in 3D.
- ai-observer: It focuses on token and cost data, traces, logs, and metrics in DuckDB, rather than a project map and 3D task view.
- OrcaReplay: It records and replays agent runs for debugging, rather than observing local files, plans, and activity through a project map.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| agenttrail | JavaScript | MIT | 913 | New |
| agent-observability | Go Template | – | 14 | Quiet |
| ai-observer | Go | MIT | 279 | Active |
| OrcaReplay | TypeScript | Apache-2.0 | 274 | New |
Considerations
The project is actively changing, with recent commits and releases, but development is concentrated in one contributor. The measured health data reports no issues closed in the recent 90-day period, so adopters should not assume a particular issue-response pace. Kitchen is explicitly an experimental preview, and local provider log formats may change.
Local observation has been exercised on macOS, while automated package and adapter checks run on Linux. Windows and live Cursor validation are pending. Remote sessions without logs on the local host are not automatically discovered. Activity history and coverage differ between Map and Kitchen, and some handoffs require explicit metadata. The project is MIT licensed; review the repository's third-party notices when building or distributing Kitchen.
Found this useful?
Share it with someone who would like agenttrail.