Your computer. Your approval.
Agents work in a folder on your disk and ask before they change anything. Here is what protects that today, what is not done yet, and how to report a flaw.
Three things that hold by default.
How the app and the API ship, before you change a single setting.
Work stays on your computer
Agents, threads, files and memory live in a folder on your disk. Nothing about your use is sent without your yes.
Every change asks first
Edits, commands, pushes and messages wait for a yes. Each decision is recorded.
The API keeps no prompts
It records each call’s model, tokens and cost. Never the prompt, never the reply.
Three places, and what each of them holds.
01
The Lobstack app
On your computer: Windows, macOS or Linux.
Each workspace is its own database file. Credentials sit beside them in an encrypted vault.
02
The Lobstack API
On Vercel, with Postgres on Supabase.
Your org, hashed keys, each call’s metadata and cost, what you share with your team, and webhook events.
03
Phone approvals
Through lobstack.ai, once you pair a phone.
Cards are sealed on your computer, and lobstack.ai can’t read them. Payments can’t be approved from a phone.
Where your data goes
- The company
- An American company, based in Chicago.
- The models
- US model companies only: Anthropic, OpenAI, Google and xAI.
- Our servers
- Vercel, in Washington, D.C. (US East).
- Our database
- Supabase, on AWS in Northern Virginia (US East).
Where a model company processes a request is its policy, not ours. Subprocessors lists each one.
How each secret is held, from your vault to our database.
- Connector credentials
- AES-256-GCM in the app’s vault. The key is kept in your system’s credential store, or in an owner-only file when that store is unavailable.
- Lobstack API keys
- Stored as a SHA-256 hash. Shown once, scoped and revocable.
- Your provider keys
- AES-256-GCM, bound to your org and to that provider. Only the last four characters are ever shown.
- Connector sign-ins
- The token passes through lobstack.ai encrypted, and is deleted once the app collects it.
- Phone answers
- Acted on only after your computer checks the phone’s signature.
Who can do what, and a record of who did it.
Roles and two-step login
Owner, admin, member and viewer; viewers change nothing. Anyone can add an authenticator-app code to their login, and on Agency and Scale an owner can require it.
The starter approves
On a shared agent server, only the person who started a run can approve it. Anyone can deny or stop it.
An audit log
Who changed keys, members, invites and webhooks. Download it as CSV, or pull it into your SIEM. On Team.
Every approval kept
In the app, each decision notes whether a person or a rule made it, and when it came from a phone.
What is in place, and what it does not stop
A page or an email can carry instructions meant for the agent reading it. The approval gate is the backstop: nothing that changes something runs without your yes.
In place
- Tool output reaches the model labelled as untrusted data, and a page can’t close the label early.
- Override phrasing, such as “ignore previous instructions”, is flagged on the step.
- A call unlike the ones you approved before is marked on its approval card.
Does not stop
- A model persuaded anyway. Labels make it less likely, not impossible.
- Paraphrase, other languages or encoded text. The flag catches formulaic attempts only.
- A harmful call to a familiar target, such as the repository an agent always pushes to.
- Labelling on the GitHub Copilot engine, which formats tool output itself.
What is not done yet, said plainly.
These are the gaps a security review would ask about. The security docs list every control and its state.
No SOC 2 audit
No audit has been performed, and there is no report to send.
No penetration test
No third party has tested Lobstack yet.
No single sign-on
There is no SAML or OIDC sign-in in the product.
One region
One database and one deployment, in the US, with no residency commitment.
Network jail on Linux only
On Linux with bubblewrap, a command in a repository with an allowlist reaches only those hosts. Elsewhere it is not fenced.
One trust level on a shared server
Everyone with access sees every thread and can use every connector.
Found a flaw? Tell us privately, on GitHub.
Reports go through GitHub’s private vulnerability reporting, not email. We’re a small team, so these are aims, not guarantees.
- In scope
- lobstack.ai, the Lobstack API, Console, the Lobstack app, and the public SDK, CLI and MCP repositories.
- Out of scope
- Denial of service, spam, social engineering, and attacks that need your computer already compromised.
- What to send
- What you found, how to reproduce it and what it lets someone do. For the API, add the call’s
x-lobstack-request-id. - Safe harbour
- Good-faith research is welcome. Test only your own accounts, keep no one else’s data, and don’t degrade the service, and we won’t take legal action over it.
- Disclosure
- Please stay quiet until a fix ships or 90 days have passed since your report, whichever comes first. We’ll agree the date with you.
- Rewards
- There is no bug bounty. With your permission, we credit you in the published advisory.
Questions, answered.
No. The app’s repository is private. What you can check instead: the installers are signed, the security docs list every control and its state, and every call comes back with a receipt. The npm packages and the developer skill files are MIT.
Your prompts, an agent’s instructions and anything a tool reads, to the provider serving that call and no other. On Copilot or your own key, they go straight to that provider.
No. Lobstack does not use your prompts, replies or other content to train models. Each model company has its own policy for what it receives.
The request id, the models asked for and served, tokens, cost, status and latency, and the key and org it was billed to. Not the prompt, and not the reply.
Request rows are kept for now. Billing records are kept for seven years, and platform logs for ninety days. The full schedule is in Privacy §8.
In the app, delete the workspace; it is a folder on your disk. For your account, ask us and we delete it within 30 days, and from backups within 90. Privacy §8 explains how to ask for deletion.
Vercel and Supabase host the site and database, Stripe takes payments, Resend sends email, Airtable holds sales enquiries, and model companies serve calls. Privacy §6 names each one.
An update check, carrying your platform, architecture and version. Nothing else without your yes: a report leaves only when you press Send, and one anonymous first-answer time only if you say Yes when the app asks, once.
Your computer. Your approval.
Report a flaw through GitHub’s private reporting, never by email. For anything else, write to hello@lobstack.ai.