What we collect

Generated from the Revealyst Agent’s on-device allowlist.

The optional Revealyst Agent summarizes your local Claude Code sessions on your machine, and the desktop app can also check which known AI desktop apps are open (their names only, from a fixed list) and, on your computer only, read your prompts to guess the kind of taskyou’re doing (from a fixed list like research, drafting, or coding). It pushes only aggregates. This page is generated directly from the agent’s field allowlist — the same list the code enforces — so it can never describe sending less, or more, than the agent actually sends. The categories below are the whole story: a small set of values that leave your machine, and a larger set read only on-device and reduced to counts or labels before anything is sent. The words you type are never among the values that leave.

Values that leave your machine

Only these leave your device, each as a number or a sanitized label — never free text:

• Model id — The model id (e.g. claude-…) is sent as a metric label, sanitized to a safe charset and length.
• Input tokens — The input-token count is summed per day and sent as a number.
• Output tokens — The output-token count is summed per day and sent as a number.
• Cache-read tokens — The cache-read-token count is summed per day and sent as a number.
• Cache-write tokens — The cache-write-token count is summed per day and sent as a number.
• AI app in use — The desktop app checks which known AI desktop apps are open (from a fixed list) and sends only each app's name as a label — never its windows, files, or anything you type in it.
• Kind of task — The desktop app reads your prompt on your computer to guess the kind of task (from a fixed list like research, drafting, or coding), and sends only that one label plus a daily count — never the words you typed.
• Refinement turns — How many of your prompts that day were follow-ups that refine an earlier answer, worked out on your computer and sent as a plain number — never the words you typed.
• Checking AI output — How many of your prompts that day asked to check the AI's work (for example verify, cite a source, or test it), worked out on your computer and sent as a plain number — never the words you typed.

Read on your machine only (never transmitted)

These are inspected locally to keep counts honest, then reduced to counts or day/hour buckets. Their values never leave your device:

• Record type — Whether a line is an assistant reply, a prompt, or activity — counted, never its text.
• Session id — Groups lines into sessions and measures overlap on your machine. The id itself never leaves.
• Timestamp — Bucketed on your machine to a calendar day and an hour-of-day histogram. The exact time never leaves — only the day and coarse hour counts do.
• Sidechain flag — Distinguishes your sessions from subagent work so session counts stay honest. A boolean, counted only.
• Request / message id — De-duplicates streamed reply lines so usage isn't double-counted. Used on-device; never transmitted.
• Line uuid — Fallback de-duplication key when no request id is present. Read on-device only.
• Content block type — Only the TYPE of a content block is inspected (e.g. tool-result vs text) — the block's contents are never read.
• Tool-result marker — Presence-only: marks a line as tool output so it isn't miscounted as a prompt. The tool output itself is never read.

The words you type in a prompt are also read on your computer only — used to count them and to guess the kind of task (from the fixed list above) — and then dropped. The prompt text itself is never stored, never copied, and never sent. Only the counts and the task label leave.

Never leaves your computer

None of the following is ever uploaded, stored, or sent. Most of these the agent never even reads; the words you type it reads only on your computer (to work out the counts and task label above) and then drops — the text never leaves. There is no content field anywhere in the Revealyst data model.

• Prompt text and assistant replies
• Tool inputs and tool outputs
• File paths and working directories
• Git branch names
• Session titles and summaries
• Queued or draft prompts

How this compares to the connectors

API-key and app connectors (Anthropic, OpenAI, Cursor, GitHub Copilot) pull the behavioral usage metrics the vendors already expose on their admin APIs — never prompt or completion content, which those APIs do not return. See the Privacy Policy for the full data-handling detail.