Skip to main content
Each developer installs the agent on their own machine. It takes about a minute and needs no administrator rights.

Before you start

  • Node.js 22 or newer. Check with node --version.
  • A nOps login for your organization. A developer without one can create it while signing in; they get the restricted Developer role. See Who sees what.
  • Your admin has set the allowed organizations in Settings → Integrations → AI ROI Agent → Policy. Without them, the agent installs but records nothing. See Manage the agent.

Install

1

Run the install command

In a terminal:
The agent prints what it will send to nOps and what stays on the machine, then opens your browser.
2

Sign in and approve the machine

Sign in with your nOps login. The approval page shows the same list of what is sent and kept. Type the code shown in your terminal, check the machine name, and click Approve.The code proves the request came from your terminal. nOps never shows it on the page, so a link someone else sends you can’t approve a machine.
3

Done

The terminal confirms the machine is signed in. The agent now runs in the background and starts with your computer.
Approving a machine is your consent to what the agent sends. Each machine gets its own key: there is no shared key, and an admin can revoke one machine without affecting others. Running the command again on a machine that is already signed in prints Already running as <your email> and changes nothing.

Check it works

  1. In a repository that belongs to one of your allowed organizations, ask Claude Code or Cursor for a small change, and commit it.
  2. Wait a minute, then run:
    Every line should say [ok]. “Local labels” may say the model downloads on first use; that’s normal.
  3. The next day, open AI ROI → Developers in nOps. Your work shows under How you use AI.
If something isn’t [ok], clara doctor says how to fix it. See Commands and troubleshooting.

What the install changes

The agent changes only the developer’s own settings.
  • AI tools: adds hooks to Claude Code, Cursor, Codex and GitHub Copilot so the agent hears about sessions, prompts and tool calls.
  • Git: adds global hooks for commits, merges and checkouts. With git 2.54 or newer they run next to each repository’s own hooks (Husky, lefthook and others keep working). With older git, the agent sets a global hooks folder that runs your existing hooks first. Git hooks never block or fail a git command.
  • Shell: adds a short block to ~/.zshrc and ~/.bashrc (PowerShell profile on Windows) so commands run inside allowed repositories are counted, by name only.
  • Claude Code telemetry: turns on Claude Code’s built-in cost reporting and points it at a relay on the machine (127.0.0.1:43180). The relay strips anything that could contain prompt, response, tool or file content before forwarding. No key is written into Claude Code’s settings.
  • Clara MCP: adds the Clara MCP server to Claude Code, Cursor and Codex. Other MCP servers are left as they are.
  • Background service: a user service (launchd on macOS, systemd on Linux) that sends data and labels prompts. On Windows there is no service: a hidden background run starts with each AI session and stops after 10 idle minutes.
  • Local folder: ~/.nops/clara, which holds the agent itself, its key, data waiting to be sent, and the labelling model once downloaded.
clara uninstall removes all of these and revokes the machine. See Uninstall.

The labelling model

The first time there is a prompt to label, the agent downloads a small model (about 310 MB) and its runtime (about 85 MB) into ~/.nops/clara. It runs at low priority, only while there are prompts waiting, and then exits, so it uses memory (about 0.65 GB) only while labelling, about 30 ms per prompt. If the model can’t run on a machine, labelling turns off and everything else keeps working. If your admin turns prompt labels off, neither is downloaded.

Windows notes

  • Git hooks run through Git for Windows.
  • Shell commands are counted through the PowerShell profile. Windows’ default execution policy skips profiles, so commands are counted only after:
    Only where your IT team allows it.

Updates

The agent keeps itself up to date. It checks for a new version each time its background service starts and about every six hours, verifies the download’s checksum, test-runs it, and only then switches over. If a new version fails to start, it goes back to the previous one. See Updates.

Next steps