Skip to main content
The AI ROI Agent is built so that prompts, AI responses and code never leave the developer’s machine. Developers see the lists below before they approve a machine, again in clara status, and admins see them in Settings → Integrations → AI ROI Agent → What’s collected.

What is sent to nOps

  • The developer’s work email and the machine’s name.
  • Which AI tool and model was used, and for Claude Code its token counts and cost.
  • How many prompts, tool calls and edits each session had, and how long it ran.
  • A work-type label for each prompt (Build, Fix, Understand, Docs, Ship, Other or Follow-up) and how sure the model is, picked by a small model on the machine when prompt labels are on.
  • Repository and branch names, commit IDs, lines changed and how many of them AI wrote (counted on the machine), and top-level folder names.
  • The kinds of files edited, like TypeScript or tests.
  • Shell command names, like git or npm, without their arguments.
  • Ticket keys, like NOPS-123, from branch names.

What stays on the machine

  • Prompts, AI responses and code.
  • File names, file contents and commit messages.
  • Screens, keystrokes, clipboard, browsing, environment variables and credentials.
  • Anything in repositories outside the GitHub organizations, GitLab groups and Bitbucket workspaces your organization allows. Personal repositories are not recorded.

How prompts are labelled

1

The prompt is queued on the machine

Prompts waiting for a label are kept, with secrets redacted, in ~/.nops/clara/pending on the developer’s machine.
2

A small model on the machine reads it

EmbeddingGemma 2 (an open model from Google, Apache 2.0 licence) turns the prompt into numbers, and a small label table on the machine picks the work type. Nothing is sent to an AI provider or to nOps to do this.
3

Only the label is sent

nOps receives the label and how sure the model is. The prompt is deleted from the machine once labelled. The model’s numbers stay on the machine too, because they could be turned back into text.
nOps may publish improved label tables. The agent downloads one only when its checksum matches what nOps announces, and keeps its built-in table otherwise. A label table contains numbers only, no prompts. Admins can turn prompt labels off for the whole organization. Then no model is downloaded and no labels are sent.

Claude Code’s own telemetry

Claude Code can report its own cost, tokens, lines, commits and PRs. The agent turns this on and sends it to a relay on the developer’s machine (127.0.0.1:43180), not straight to nOps. The relay:
  • drops every log body, and every field that could hold prompt, response, tool or file content, even if Claude Code’s own content logging is turned on;
  • forwards what is left to nOps with the machine’s own key.
No key is written into Claude Code’s settings.

Which repositories are recorded

The agent records only inside git repositories whose remote belongs to an allowed organization: A fork counts when its upstream remote belongs to an allowed organization. Everything else, including personal repositories and open-source projects outside your organizations, is ignored.

How data travels and is stored

  • On the machine: the machine’s own key and data waiting to be sent live in ~/.nops/clara. The key is used only to send telemetry and fetch the organization’s policy.
  • In transit: HTTPS to api.nops.io, every 30 seconds. While offline, data waits on disk.
  • At nOps: each event is stamped with the machine’s identity from its key, so one machine can’t send data as another person.
  • Who sees it: see Who sees what.

Developer controls

See Commands and troubleshooting.