Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

Cloakroom: hide your secrets from your agent

Written November 11, 2026

Cloakroom sits between your coding agent and your files, swapping keys and personal details for placeholders before the model sees them.

The problem

Point a coding agent at a project and it reads everything it can reach: the .env file, the output of the last command, a support ticket with a customer's name and email, a log with a live API key. All of that goes to the model. Once a key is in the model's context, you cannot take it back, and rotating it after the fact is the only repair.

How it works

Cloakroom runs as hooks around the agent's tool calls. On the way in, pattern rules plus a small local detection model find secrets and personal data, and each value is swapped for a placeholder such as PII_SECRET_1 that stays the same for the session. The pairs live in an encrypted vault on your machine, with its key in the operating system keystore, and the agent cannot read the key. On the way out, placeholders in file writes and commands get their real values back just before they run, so the files on your disk still hold the real values.

Everything runs locally. The only network call is a one-time download of the detection model. If Cloakroom cannot check something, it holds it instead of letting it through. A command that would send a known secret off your machine is held until you say so, and a key written into a file git would commit is refused.

What it covers

Cloakroom works with Claude Code, Codex CLI, Gemini CLI, and Cursor's command-line agent, and masks API keys, passwords, emails, names, addresses, and card and bank numbers. The current release is 1.4.0, which extends the exposure check to personal data typed in old chats across Codex, Cursor, and Gemini CLI, not only Claude Code. Coverage is not identical across agents: the README carries a table of what is masked, held, or not covered per agent, and the gaps matter. Output from a long-running command on Codex reaches the model unmasked, and Gemini CLI lets a call through if a check runs past its time limit.

How to install it

In Claude Code, one line at a time, then restart it:

/plugin marketplace add shanjeevrajendran/cloakroom
/plugin install cloakroom@cloakroom

For Codex, Gemini CLI, or Cursor, on Mac or Linux:

uv tool install "cloakroom[model,torch,ocr]"   # Python 3.10+
cloakroom install codex                         # or: gemini, cursor

Codex skips new hooks until you trust them once, from /hooks inside Codex. Then run cloakroom doctor. Every line should say ok.

The commands you will use

CommandWhat it does
cloakroom doctorHealth check, versions, per-check timing
cloakroom exposureLists the keys and passwords your coding agents saw, so you know what to rotate. Hints only, never the key itself
cloakroom scanShows the commands a repo's agent and editor config would run
cloakroom check FILEPreviews what would be masked in a file, and why
cloakroom dashboardAlerts in the browser. Mark a false alarm once and a rule stops it next time
cloakroom updateUpdates the program and sets the hooks up again

The exposure list is hints only. It tells you which keys an agent saw so you can rotate them. It never shows the key itself.

Results, as reported

On held-out test sets, Cloakroom reports letting through 0 of 1,005 personal details and 0 of 77 keys and passwords in real vendor formats. The other tools it was tested against let through 119 to 992. Those numbers are the project's own, on its own test data, and no independent test exists. The trade-off is false alarms: it masks more harmless text than secret-only scanners, and false alarms are its main weakness.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Secret scanning now catches Actions log leaks · Lazarus: ten neurons can backdoor an agent · Clean data can strengthen a hidden backdoor