Omna on your computer

AI agents

AI agents like Claude Code read your files, your terminal output and your configs — exactly where keys and customer data live — and send them to an AI on their own. Omna masks them before each request leaves your computer, puts personal data back in what the agent writes so your code still works, and leaves a receipt for every request.

3 min readClaude Code, Codex, Cursor and your own agents

Which agents are covered

AgentHow
Claude CodeConnected automatically by the install command
aider, Codex CLI (also when signed in with a ChatGPT account), VS Code (Copilot Chat and similar), ContinueConnected automatically if already installed. Installed one later? omna enable aider (or codex, continue)
Cursor, Cline, WindsurfOne setting, typed once inside the tool — see Supported tools
Your own app using the Anthropic or OpenAI SDKPoint it at Omna: ANTHROPIC_BASE_URL=http://127.0.0.1:7788 or OPENAI_BASE_URL=http://127.0.0.1:7788/v1

What the agent sees, and what your files get

  • The agent sees only masked data. A key in a file it reads becomes [SECRET_AWS_KEY_1]; an email becomes [EMAIL_1].
  • Personal data comes back; secrets don't. An email or a name is put back in what the agent writes. A key is not: a poisoned web page or file could talk the agent into a command that sends [SECRET_GITHUB_TOKEN_1] somewhere, and putting the real key in would hand it to a stranger. The cost: an edit to a line holding a key won't match your file, and a command that needs the key fails visibly — use the environment variable ($GITHUB_TOKEN) instead. Accept the risk? omna tool-calls secrets on.
  • Always labels, never fake look-alikes. For coding agents, Omna always uses labels like [EMAIL_1]. A leftover label is obviously wrong and gets noticed; a leftover fake email looks real and could be committed by mistake. So the realistic style is refused for anything that writes files.

Want secrets removed for good, not put back? Start Omna with omna start -d --no-restore-secrets.


Your own agent

If your agent uses the Anthropic or OpenAI SDK, point it at Omna. Your key, your model and your code stay the same.

sh
export ANTHROPIC_BASE_URL=http://127.0.0.1:7788   # Anthropic SDK
export OPENAI_BASE_URL=http://127.0.0.1:7788/v1   # OpenAI SDK

Or set it in code:

python
import anthropic
client = anthropic.Anthropic(base_url="http://127.0.0.1:7788")

from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:7788/v1")

Real values come back in streamed answers. Personal data also comes back inside tool-call arguments, so your agent's tools receive the real data the AI never saw; secrets don't, unless you run omna tool-calls secrets on.


The proof

Every agent request leaves a receipt: when, which AI, how long hiding took, and what was hidden — counted, never shown. omna report turns them into a summary for this computer; for a whole team, see the sample report. It lists each agent and tool your team ran and what was caught.


If Omna isn't running

Your coding agents can't connect at all — nothing goes out unhidden. Start it again with omna start -d. Claude Code also restarts it by itself at the start of every session.


Check it's working

sh
omna status    # running? which tools are connected? requests today
omna log       # one line per request, with what was hidden

Good to know

  • A tool's own cloud features (for example Cursor's tab-completion) don't pass through your computer, so Omna can't see them.
  • An agent's "thinking" is passed through untouched (the AI companies require it), so you may see a masked label there.

What's next