Skip to main content

Security overview

Your raw files are never uploaded to Locus. Depending on the features you use, chunk text (managed embeddings) and search excerpts with file paths (remote connection) pass through our infrastructure in transit, and we don't store them. Account data is stored. What follows is what protects that surface today. The full breakdown by mode is on /privacy.

Last updated 2026-10-07.

How data flows, by mode

Your raw files are never uploaded to Locus.

  • Local engine, local embedder. Indexing and embeddings run on your computer. When your AI assistant searches, the matching excerpts go to that assistant, as they would if you pasted them in.
  • Your own embedding key. Chunk text goes straight from your computer to the embedding provider you chose, under your own account.
  • Managed embeddings. On managed plans, chunk text and your search queries pass through Locus's relay to Voyage AI to be embedded. We don't store them; we keep only usage counts.
  • Remote connection. When you reach your computer from another device or a web AI app, your question and the matching excerpts travel through Locus's relay (Vercel and Supabase) to and from your computer. They are encrypted in transit and we don't store them.
  • Account. Your account details are stored with Supabase. Your files are not.

The full table, including what each mode sends to which vendor, is on /privacy. Our vendors are listed on /subprocessors.

What we do today

  • Our website and API are served over HTTPS by our hosting platforms.
  • Every account-data table in our database has Row Level Security enabled (details below).
  • Our continuous-integration pipeline scans every commit for committed secrets with gitleaks and runs Semgrep static analysis.
  • We ran dependency vulnerability audits once, on 2026-09-23, and track what they found. There is no scheduled dependency scan yet.
  • API keys are stored as hashes, and you can revoke your own keys at any time (details below).
  • There is a written vulnerability-disclosure policy with a contact address (see "Reporting a vulnerability").

Built, but not counted as protection today: signing of the requests the relay sends to your computer, and automatic spend stops on managed embeddings, exist in our code. They are not switched on in production, so we do not rely on them or ask you to.

Data isolation — Row Level Security

Every account-data table has Row Level Security enabled in our Postgres database. Your account can read only its own rows (plus the organization records described below, if you join an organization) and cannot write to the API-key table directly: keys are created and revoked only through database functions that check who is asking and read your plan's key limit from your subscription, not from the request. Our server-side routes use a privileged database key to check API keys and record usage; that key is never sent to your browser.

If you join an organization, its members can see your email, your role, when you joined, whether you are still a member, when you last used Locus's hosted features and which Locus version you ran. Its admins can also see its invites and audit log, and can revoke your Locus API keys. Nobody in the organization can see your files or searches, because Locus doesn't store them. As the operator, we can read account records in the database for support and billing.

API keys

API keys are stored as hashes, never in plaintext. You can revoke your own keys from the dashboard at any time, with no admin approval needed. The embedding relay, the remote-connection relay and the account-status check look the key up on every request with no caching, so a revoked key is refused on its next request. One exception: a relay token already created from the key stays valid for up to ten minutes. A relay token is an account-level session, not only a connection for your computer: for its lifetime it can act as your account. Our database refuses relay tokens for creating or revoking API keys; they are not otherwise limited to the relay yet. Whether a relay connection that is already open is closed the moment its token expires is not yet verified.

Anyone who holds your Locus API key can search and read your indexed files through the relay while your computer is serving, take over as your relay computer, spend your managed-embedding budget, and act as your account for up to ten minutes with each relay token they get from it. Keep it secret. If it leaks, revoke it from the dashboard, then check your key list again after ten minutes and revoke any key you did not create.

Local data directory

The local engine — the part that actually holds your indexed chunk text and vectors — stores its data in ~/.locus (or a directory you configure via LOCUS_DATA_DIR). That directory is created with 0700 permissions: owner-only read, write, and execute. Other local OS accounts on the same machine can't read it. This limits access on a shared machine; it does not encrypt the data at rest — full-disk encryption (standard on most modern laptops) is the real first line of defense for a lost or stolen device, and Locus relies on that rather than re-implementing its own at-rest encryption.

Locus also refuses to follow symlinks during indexing or file reads, so a symlink placed inside a folder you point Locus at can't be used to reach content outside that folder. locus_read_file (the MCP tool an AI assistant calls) can only read a path already present in your local index — never arbitrary filesystem access.

Multi-factor authentication

We do not claim that multi-factor authentication is enforced on every account that controls our production infrastructure.

What we do not have

  • No SOC 2 report exists, because no audit has been performed.
  • No ISO 27001 certificate exists.
  • No penetration test has been run.
  • No HECVAT has been submitted.
  • No dedicated security team or on-call rotation: one person handles security operations today.
  • No log-retention or SIEM tooling of our own; visibility into what happened relies on each vendor's own dashboard and audit log.
  • No incident-response drill has been run.

SOC 2: aligned, self-assessed, not audited

Aligned with the SOC 2 Trust Services Criteria: self-assessed, not audited or certified.

We keep a self-assessed control map against the SOC 2 Trust Services Criteria (Security; Availability and Confidentiality as optional add-ons); no auditor has reviewed it and no SOC 2 report exists.

Of 39 mapped rows, 11 are implemented, 23 partial, 4 missing and 1 not applicable (2026-10-07); multi-factor authentication on vendor accounts is not yet confirmed.

SOC 2 is on our roadmap; no audit has been engaged.

Roadmap (planned, not promised)

These are plans, dated by quarter. They can slip, and none counts as done until this page says so.

  • Q4 2026 (planned): confirm multi-factor authentication on our vendor accounts, turn on and test database backups and a restore, sign an access review, approve our written security policies, add scheduled dependency scanning.
  • Q4 2026 to Q1 2027 (planned): alerting, a recorded incident drill, vendor security reviews, and a decision on a compliance platform and an auditor.
  • 2027 (planned, no date): an independent audit could follow only after the above and only if it is funded. We have not engaged an auditor and make no commitment about when, or whether, a report will exist.

Reporting a vulnerability

If you find a security issue in Locus (the local engine, the MCP server or this website), email yonilev2003@gmail.com with what you found, the steps to reproduce it and its impact. Please do not open a public issue first. Our policy is in SECURITY.md at the root of the repository and is available on request. We aim to acknowledge a report within about 3 business days; that is the intention of a one-person operation, not a service-level agreement.

There is no bug-bounty program.

Questions from your security team

Our security overview and questionnaire answers are available on request. Write to yonilev2003@gmail.com. Answers say what is partial or missing rather than rounding up.

Living document, current-state, not a certification. Updated when our actual security posture changes, not on a fixed schedule.