How Proxon collects data

Proxon measures how AI is used by collecting it from the places it happens. Some sources watch where your people work, and others pull from your vendors or your own agents. The more you connect, the more Proxon can tell you: broader coverage of how AI is used, and sharper detail on what it costs. You do not need everything to begin, and you do not need all of it ever. Pick the sources that fit your organization, and add more when you want more coverage.

The sources fall into two groups. The recommended set covers most organizations. The developer options are for engineering teams that want to instrument their own tools and agents.

Recommended

These cover most organizations, and most rollouts start here.

Client-side observers

Client-side observers run on employee machines and give the broadest view of how AI is used. There are three, and each one watches a different place, so each catches usage the other two cannot. You install them by deploying to devices, and what they record depends on the capture level, which is explained in What Proxon captures.

• The Browser Observer sees AI used inside the browser, such as which web assistants people use and how actively. It is the only source that sees browser-based AI at all.
• The Desktop Observer sees which app is in the foreground across every desktop app, so it catches AI tools outside the browser like Cursor and terminal tools, and it shows workflow patterns such as moving an answer into a document. It is the only source that sees AI used in desktop applications.
• The Network Observer sees the calls that CLI and desktop AI tools make, including prompts, replies, tool calls, and the token counts that make per-request cost possible.

AI vendor integrations

AI vendor integrations pull usage and cost straight from the AI vendors you already pay for, without touching anyone's machine. Because the numbers come from the vendor's own billing, they are the most accurate cost source, and they capture usage that never runs on an employee's machine, such as server-side or automated calls. An admin connects a vendor admin or billing key, and Proxon pulls from there. See AI vendor integrations for the vendors and setup.

Options for developers

These cover the AI your engineering teams run themselves. Use the Capture Proxy to bring AI coding tools like Claude Code and Codex into Proxon. It takes a one-line change to the tool and no code. Use the Proxon SDK for Agents or OpenTelemetry for the agents and services you build in-house. These often run server-side or in CI, where the client-side observers cannot see them, so instrument them to make sure Proxon does.

Capture Proxy

The Capture Proxy brings AI coding tools like Claude Code and Codex into Proxon. The developer changes one base URL in their tool, and its prompts, replies, and cost start showing up.

Proxon SDK for Agents

The Proxon SDK for Agents adds per-agent cost and token telemetry to a Node app with one launch flag and no call-site changes. It reports economics per agent and task, and never sends prompt or reply content.

OpenTelemetry

OpenTelemetry reads usage and cost from anything that already emits it, in any language, including server-side AI and tools that already speak OpenTelemetry. There is no new agent, proxy, or certificate.

How much each source sees

The three observers share a single capture level, which an admin sets. Behavioral data is the default, and the most detailed level, Content aware, stays off until an admin turns it on. The developer options carry only what they are given: the Proxon SDK for Agents never sends prompt or reply content, OpenTelemetry sends only what your exporter includes, and AI vendor integrations pull vendor usage and cost rather than content. See What Proxon captures for what each level includes.

Choosing your sources

Most teams start with the recommended set, since it installs through the tools you already use and gives broad coverage on its own. Add a developer option when you want to see your own agents and backend services, or an AI vendor integration when you want vendor-grade cost accuracy. You are never locked in, since an admin controls every source and its capture level at any time.