What leaves your laptop
Metadata about each run: what was asked, which tools ran on which paths, and how it ended.
- Prompt
- What you typed, with the agent’s system reminders removed and secrets redacted. The first 1,000 characters are stored; the first 300 are sent live for recall.
- Tool calls
- The tool’s name and its file paths, search patterns, shell commands (first 1,000 characters, redacted), URLs and web queries, and the tool’s short description.
- Where it ran
- Machine name and folder path; at session start the project name, git branch, model name and the path of the agent’s transcript file (not its contents).
- Done-checks
- The check commands the agent ran, whether they passed, and how the turn ended.
- Files touched
- Git hashes of the files a turn read or wrote, so stale steps can be spotted.
- Token counts
- Tokens used per turn, read from the agent’s transcript on your laptop.
- What ArcRoom handed over
- The size of each piece of context ArcRoom added, and the text of hand-overs (up to 2,000 characters).
- What agents save on purpose
- Whatever an agent sends to ArcRoom’s own tools (remember, complete_task and the others): their input and output.
- Failed commands
- When a command fails: the command and the first 500 characters of its error (redacted), and what worked instead, so the next agent is warned before it repeats it.
- Code map
- Your repo’s folder and file paths with their git hashes, and which files your agent is editing right now (with your name and branch). Shared with your workspace. On by default. arcroom code off stops it for a repo.
- File contents, whether read or written
- Tool output, except the error text of a failed command
- Edit diffs
- Prompts given to sub-agents
- Your AI provider account and its keys
Two exceptions: when your agent saves something to team memory with ArcRoom's tools (remember, complete_task), that text is sent, even if it quotes a file. Prompts are sent with whatever you paste into them. Every saved memory is listed on the Memory page, so you can see what was kept, and arcroom off stops ArcRoom for a repo on your laptop.
The plugin never sends what is inside .env or key files. A path or command that names one is sent as [secret file], and secret-looking values from your project's .env are removed wherever they appear (best effort, see Redaction). Details in What is sent.
Redaction
Secrets are removed on your laptop before an event is queued, again before it is uploaded, and from every request the plugin sends. It removes:
- This machine’s secrets
- The exact values of your secret-named environment variables (…_KEY, …_TOKEN, …_SECRET, passwords) and the secret-looking values in the project’s .env files, wherever they appear and whatever their shape.
- Known key formats
- Anthropic, OpenAI, GitHub, GitLab, AWS, Google, Stripe, Slack, npm, Hugging Face, Shopify, Twilio, SendGrid, Mailgun, DigitalOcean, Linear, PyPI, Doppler, Postman and Databricks keys, JWTs and private key blocks.
- Secrets by name
- name=value and "name": "value" pairs whose name says password, secret, token or key, and any field under such a name.
- In URLs and headers
- Passwords inside URLs (https://user:password@host), Bearer and Basic auth headers, cookies, signed-URL signatures.
- On commands
- Secret flags such as --password, --token, --api-key, mysql -p and curl -u user:password.
It is best effort. A secret in a shape none of these match can still get through. Please don't paste secrets into prompts.
Safety on your laptop
If ArcRoom fails, your agent carries on without it: memory lookup gives up after 3 seconds and the prompt goes through. ArcRoom stops an action on purpose only in three cases:
- Loop guard: the third identical tool call in a row is blocked.
- Done-checks: if your approved checks haven't passed since the agent's last edit, it is asked to run them before it stops. It is sent back at most twice, then let go and the run is marked partial.
- Same file: if another agent is already editing a file, your agent's first edit to it is held back once and told who is on it. If your agent makes the edit again, it goes through. If ArcRoom can't check in time, the edit goes through.
- Never stalls a session when something breaksAn error inside ArcRoom lets your agent’s call through, and ArcRoom always exits cleanly.
- Memory lookup gives up after 3 sThe prompt goes on without memory and you see
memory offline. Lookups then pause for 60 seconds, so an outage costs that wait once a minute, not on every prompt. - Uploads happen elsewhereArcRoom only adds events to a local queue,
~/.arcroom/spool.jsonl, and a separate background process sends them within seconds. Offline, events wait in the queue and are sent when you're back online.arcroom offstops a repo, including anything still queued for it. - Off per repo
arcroom offstops recording and sending for that repo on your laptop.arcroom onturns it back on. - Done-checks: you approve, the agent runsArcRoom never runs them itself. The agent runs them with its own shell, under its own permission prompt, and only after you approved those exact commands with a code (
arcroom checks approve <code>) in your own terminal.
Details in Fail-open behaviour and Done-checks.
Controls you have
arcroom off- Nothing from a repo is recorded or sent from your laptop. The setting lives on your laptop, not in the repo.
arcroom sent- What this laptop sent, day by day, for the last 7 days (--all shows every event exactly as sent).
- Yours or the team’s
- Decisions your agent saves stay yours unless it saves them for the team or you share them. Finished work, the answers and steps your agents complete, is shared with your team so nobody redoes it; ask your agent to keep a result private before it starts the task, and only you can see or reuse that answer. To share something later, press Share with team on the Memory page (or run arcroom share in the project).
- Revoke a key
- In Settings → API keys. ArcRoom stops on that laptop by itself.
- Delete a project or workspace
- Removed from ArcRoom immediately, with no trash and no restore. We keep no backups, so they can’t be recovered.
arcroom wrong- Report a hand-over that was wrong. Wrong answers are thrown away and broken steps fall back to their last good version.
Every hand-over carries its provenance: steps say who saved them, when, and whether they passed their checks.
How long we keep it
- Run events
- The step-by-step trace. Deleted after your plan's history: 7 days on Free, 90 days on Pro.
- Run summaries
- Prompt, outcome, token counts and report. Deleted with the run events, after your plan's history.
- Team memory
- Answers, steps, decisions, failures and rules. Kept while the workspace exists; a failure no agent hits for 30 days is removed.
- Answers
- Handed over while fresh: 1 day by default (owners can set 1 hour to 30 days), or the lifetime the agent gave it, up to a year; each reuse restarts it. A pinned answer doesn't expire. Expired answers stay stored with their steps until you delete them. Details in Answers, steps and expiry.
- Sign-in sessions
- 30 days, or until you sign out.
- Deleted projects and workspaces
- Removed at once. We keep no backups, so they can’t be recovered.
How your data is protected
- Through the API only
- Every read and write goes through the ArcRoom API, which checks your key or session. The database cannot be reached from the internet directly.
- Encrypted in transit
- All traffic to ArcRoom uses HTTPS.
- No secrets stored in plain text
- Passwords, API keys and sign-in tokens are stored as hashes.
- Matching by meaning
- The first 300 characters of a prompt (already redacted) and saved answers and decisions are turned into embeddings by Google Vertex AI so similar tasks can be matched.
We never use your data to train a model.
The providers that run ArcRoom for us are listed in the privacy policy.
Reporting a problem
Found a vulnerability? Write to sachin.kumar@arcroomai.com. Please don't test against other people's workspaces.
Not yet
ArcRoom has no self-hosted version, no SSO and no SOC 2 report today. We'll start each one when the first team that needs it asks. No dates until then.
For what we do with personal data, see the privacy policy.