The more Proxon observes, the more it can tell you.

Proxon reads AI usage from a few kinds of source. Some watch where your people work, others pull from your vendors or your own agents. The more you connect, the stronger the signal: broader coverage of how AI is being used, and sharper detail on what it costs. You don’t need everything to begin — and you don’t need all of it ever. Browser-only is a supported configuration. Add the Network Observer or a Direct Integration when you want per-request token counts and cost. Pick the sources that fit your organization and add more when you want more coverage.

You stay in control · for IT & Security

Every observer can be turned on or off and set to one of three capture levels. Each level includes everything above it, and you set the level per observer. Behavioral is the default; content stays off unless an admin turns it on.

Metadata. The facts of an interaction: which AI tool, when, and how much (token counts and cost), without any of the words. It’s what drives adoption and most cost signals.
Behavioral. Adds patterns of use: engagement signals like active tabs, typing, and app focus that show how AI fits into real work, still without the words themselves. This is the default level.
Content. Adds the words: prompt text and model responses, plus the names of documents, pages, and windows people open. Turning it on is a single, explicit choice, and what each observer captures at this level is spelled out on its card below.

Throughout, “typing” means how much typing is happening — volume and rhythm — never the keystrokes themselves. No observer records what you type.

Admins manage all of this in Data Collection under org settings and can change any observer’s level at any time.

Core Observers

Although you can choose which observers to use, we recommend all three for most Proxon rollouts. Each one watches a different surface where your people use AI — the browser, the desktop, and the network — and each catches usage the other two can't see, so the more you turn on, the broader the picture. A browser-only deployment is a supported configuration: pick what fits your organization, and change your mind at any time.

Browser Observers

Optional

extensions for Chrome, Edge & Firefox

Uniquely collects: AI used inside the browser: which web assistants (ChatGPT, Claude, Gemini, and others) people are on, plus engagement signals like which tab is active and how actively it is being used. The only mechanism that sees browser-based AI at all.
How it works: Runs as a browser extension, observing AI tab activity locally in the browser and reporting metadata by default. Below the content level it records a site’s domain, not the full URL path or the tab title. At the content level it can also retain the full path and tab title, and on supported AI chat sites, the prompt text.

Desktop Observer

Optional

lightweight agent, macOS & Windows

Uniquely collects: Activity and context across everydesktop app: which app is in the foreground (and, at the content level, that window’s title), plus how actively the device is being used. Reveals workflow patterns like moving an answer from an AI app into a document. The only source that captures what’s happening around AI usage, including in non-AI apps.
How it works: A lightweight agent that runs on the device and observes which app is in the foreground plus engagement signals. At the content level it also records the foreground window’s title, which can include a document’s name.

What this observer never captures.It does not record your keystrokes, your clipboard contents, or the commands you run. It sees which application is in the foreground and when — and, only if your admins turn on the content level, that window’s title.

Network Observer

Optional

on-device capture for CLI & desktop apps

Uniquely collects: The content of AI interactions (prompt text, model responses, and tool-use calls), plus token counts for supported apps. Those token counts make granular, per-request cost attribution possible for CLI and desktop AI (supported apps only; some don’t expose usage).
How it works: Inspects calls to a defined allowlist of known AI hosts at the capture level your admins set; everything off the list passes through untouched: for that traffic Proxon sees the connection exists, not its contents. Apps that pin their certificates can’t be inspected. Supports macOS and Windows.

Targeted Sources

Layer these on for more detail where it matters, or to reach AI that never touches a person's device. Some sharpen what the core already sees (like vendor-grade cost data), while others fill gaps the core can't, such as your own agents and backend services. Add the ones that match your stack.

Direct Integrations

Optional

API-key links to your vendor accounts

Uniquely collects: Authoritative usage and spend straight from the vendor, the billing source of truth for the accounts you connect. Tightens cost accuracy beyond what’s inferred on-device, and captures usage that never runs on an employee’s machine, like server-side or automated calls.
How it works: An admin connects a vendor admin or billing key; Proxon pulls directly from the vendor. No device agent involved.

Proxon SDK for Agents

Optional

a drop-in SDK for your own agents

Uniquely collects: Per-agent, per-task economics from inside the agents you build: exact cost and token usage tied to a specific agent and task, never prompt or response content.Sees autonomous, programmatic activity the device observers (which watch people) can’t.
How it works: A Node / TypeScript SDK: add one line and it auto-instruments your Anthropic and OpenAI clients with no call-site changes, then reports directly to Proxon. Runs anywhere Node does: CI, containers, servers, or a laptop. A Python SDK is planned; until then, non-Node apps use OpenTelemetry.

Set it up →

OpenTelemetry

Optional

point your existing OTLP exporter at Proxon

Uniquely collects: AI telemetry you’re already emitting: structured usage and cost (model, tokens, spend). You control what your exporter sends — Proxon only needs those usage attributes. Reaches any language and tools that already speak OpenTelemetry, including Claude Code and GitHub Copilot, plus server-side AI that doesn’t run on anyone’s device.
How it works: Point an existing OpenTelemetry exporter at Proxon’s OTLP endpoint using standard settings: no new agent, no proxy, no certificate. Accepts both an agent’s native OTLP and stock OpenTelemetry spans from any language.

The Proxon SDK and OpenTelemetry do the same job two ways. For a Node app you control, the SDK auto-instruments it with one line. For anything else (another language, or a tool that already emits OpenTelemetry), use OpenTelemetry. Either is enough; running both adds detail, not new coverage.

Set it up →

Where to start

Most teams start with the Core Observers: they install through the tools you already use and give broad coverage on their own. Add Targeted Sourceswhen you need vendor-grade cost accuracy, or want to see your own agents and backend services. You’re never locked in: admins control every source and its capture level, at any time.