Is there a free, open-source way to track AI agents on a website?

What open source means here

The application's source code is public on GitHub under the AGPL-3.0-only license. You can inspect the browser snippet, event ingestion, bot classification, log import and dashboard code. You can also run a copy on infrastructure you control. The license carries obligations when you offer a modified version as a network service; read the license text for the exact terms and seek qualified advice if you are planning a redistributed product.

Open source is useful for a measurement product because the data path matters. You can check which fields the snippet sends, how crawler verification works and how long raw events remain. That is stronger evidence than a privacy slogan alone. It does not mean every deployment is automatically compliant with every local rule: configuration, notices, server access and operational practices remain the deployer's responsibility.

A public repository also lets a team test an update before deploying it. The trade-off is that self-hosters become responsible for updates, backups, TLS, access control and monitoring. If nobody on the team owns those tasks, a hosted account may be simpler even when the software itself has no license fee.

What the hosted Free plan includes

The current Free plan allows one site, 10,000 agent events in a calendar month and 30 days of dashboard history. Ordinary human page views do not count against that agent-event limit. Raw event retention is separate from the dashboard window; consult the current documentation and account screen before making a long-term plan. Paid Pro and Agency tiers exist for more sites, volume and history. Do not describe the whole hosted service as permanently free or without limits.

A small site can use the Free plan to learn whether assistants send identifiable referrals, which named bots reach its pages, and whether on-page tools are called. Crawler measurement needs a log import in addition to the script because many bots never execute JavaScript. The setup should be tested on the site's real traffic before you use a chart in a public report. The complete measurement guide walks through each source.

When you evaluate the free tier, estimate the number of agent events rather than total visits. A documentation site heavily crawled by bots may have a different event profile from a small brochure site. Decide what dashboard history your team actually needs. If your reports require a full year in the application, compare that requirement with the current plan limits before starting.

What self-hosting involves

The repository includes a Docker setup and instructions for running the Next.js application with a SQLite data file. Put a TLS-terminating reverse proxy in front of it, set the site origin and mail configuration, and give the data file persistent storage. Review the environment example rather than copying the public service's legal identity or contact details into your own installation. Then add your site's domain, install the snippet and verify the first event.

A production setup needs routine backups of the SQLite file, access to logs for crawler import, a way to deliver login email and a repeatable update process. Test a restore, not only a backup job. Limit access to the application and its data file. These are ordinary server responsibilities, but they represent real labor and hosting cost. The cloud service takes on much of that operation in exchange for its plan limits and terms.

If your site sits behind a CDN, remember that an origin access log may omit requests answered from the CDN cache or blocked at the edge. Importing that log still measures what reached your server, which may be enough for your question. For an edge-level count, use the CDN's own metrics as well. Do not compare the two totals without explaining their different coverage.

The data you can actually measure

The on-page snippet records identifiable assistant referrals, supported browser tool calls and marked goals. A server log import adds bot requests and verification where operators publish address ranges. These are separate categories. Neither a request nor a referral reveals a person's prompt; a bot request does not prove a citation. Keep the labels precise in any exported report. Plausible's AI Assistants channel already handles the referral layer for many sites, while a bot audit at the edge can handle crawler activity. Choose Agent Tracking when joining those layers with tool and goal data is useful.

The product records the page path rather than a full query string and aims to avoid storing visitor network addresses. Verify those claims against the code and your own deployment configuration before putting them in a privacy notice. A self-hosted copy also changes who operates the server and who has access to the data. The privacy page describes the hosted service, not automatically every fork or private installation.

If all you need is ChatGPT-referred human visits, your existing web analytics may already answer the question. If you need crawler counts only, a log parser can suffice. The value of a dedicated agent analytics installation grows when the team needs the sequence from discovery through on-page action and a completed goal.

A decision checklist

Use the hosted Free plan when you want to test one site quickly and its limits fit your traffic. Self-host when infrastructure control and inspectable operation are essential and you have someone to maintain it. Use a general web analytics tool beside either deployment for broad human visitor reporting. If you already use Cloudflare, review AI Crawl Control for edge requests and rules; it serves a different position in the stack.

Before choosing, answer five questions: Do you need verified crawler requests or only assistant referrals? Do you expose browser tools? What event volume and reporting window do you need? Where must the data live? Who maintains the application? Then run a one-week trial with a clear success criterion, such as identifying one crawler discrepancy or one failed tool call. A small, testable goal is more valuable than installing a dashboard and never acting on it.

The software being open source is an invitation to verify the claims. The Free plan is an invitation to measure a real site without a payment card during the pilot. Neither is a substitute for deciding which question the measurement must answer.

In short