Security and privacy
Financial data deserves care. Here is how the software handles it today, what the hosted service is designed to add, and what is not built yet.
Self-hosted: your machine, your data
One SQLite file on a computer you control. No telemetry, no accounts to link, no bank connections. Every measure in the first table below is in the code today.
Hosted: designed, not running
The hosted service is not open yet. Everything in the second table is a design for it and is marked Planned, even where some of the code already exists in the repository.
What the software does today.
Taken from the code and its tests, not from intentions.
| Measure | What it means | Status |
|---|---|---|
| Your machine, one file | Your data is one SQLite file on your computer or Docker volume. Nothing is sent anywhere. | Available |
| No telemetry | No usage reports, error reports, version checks or analytics. The only network calls the server makes are webhooks you configure, blocked for private addresses unless you allow them. | Available |
| One optional CDN request | In the browser, the optional API reference page at /docs loads its viewer script from the jsDelivr CDN. The web app itself contacts no third party. | Available |
| Hashed credentials | Passphrases are hashed with scrypt. API keys and sessions are stored only as hashes, and keys are shown once. | Available |
| Sessions and CSRF | Session cookies are HttpOnly and SameSite=Strict, with a CSRF token and an origin check on writes. Failed sign-ins and setup-token guesses are throttled, including parallel bursts. | Available |
| Scoped keys and roles | API keys carry read, simulate or write scope, capped by the person's role. Every route declares its scope, and tests walk the whole API to check it. | Available |
| Safe AI access | MCP write tools are off by default, plan edits are proposals you approve, there is no delete tool, and every apply is attributed. | Available |
| Webhook protection | Webhook destinations are checked at registration and again when connecting, so a public name that points at private space is refused. Deliveries are signed. | Available |
| Audit log | Every write by a person, key or connected app, and every refused write, is recorded for the owner to read. | Available |
| Clean errors | Unexpected failures return "internal error" with no stack trace, SQL or file path. Queries are parameterised, and requests that address __proto__ or constructor are refused. | Available |
| Account export | Export a household as JSON, or take a consistent whole-database backup while the server runs. | Available |
What the hosted service is designed to do.
Each measure is marked exactly as true today. None of the hosted service is live.
| Measure | The design | Status |
|---|---|---|
| Per-customer database files | Each customer workspace gets its own SQLite file, with no shared tables between customers, so isolation is structural, one customer's restore never touches another's, and export is "copy the file". | Planned |
| Encryption in transit | TLS everywhere, with HSTS. | Planned |
| Encryption at rest | An encrypted data volume, and backups encrypted with a separate key for each workspace. | Planned |
| Not end-to-end encrypted | Our servers can read your plans. That is what lets scheduled check-ins and AI assistants work. We will not claim otherwise. | Planned |
| No financial values in logs | A rule, enforced by a scrubbing logger that drops any field not on an allow-list and by a test that fails if a sample amount, plan name or token appears in log output. | Planned |
| Two-factor sign-in | Authenticator-app codes (TOTP) with single-use recovery codes. Owners are prompted to turn it on, not forced. | Planned |
| Automatic backups | Changed workspaces are backed up every 15 minutes, kept for 48 hours, then one a day for 30 days, with restore tests. Backups are encrypted per workspace. | Planned |
| Export and deletion | Export a whole workspace as a database you could run yourself. Deleting a workspace removes its live database at once and destroys the key to its backups; backup files are removed within 30 days and provider snapshots expire within 7. | Planned |
| EU hosting | The chosen host is DigitalOcean in Frankfurt, Germany, a single EU region. Nothing has been deployed there yet. | Planned |
| Operator access limits | Staff cannot see inside a workspace. The one exception is a read-only, time-limited support grant that the owner creates, which is audited and emailed to the owner when first used. | Planned |
The hosted service will use a few outside services, for example a payment processor, an email provider and a bot check at sign-up. They are listed on the sub-processors page and in theprivacy policy, which take effect when the hosted service opens and have not been reviewed by a lawyer. Check-in emails are designed to carry a link, not your figures.
Things we do not do, and will not.
No bank connections
No aggregation, no stored bank credentials, no account linking. You enter balances or import files.
No selling or sharing your data
Your plans are not sold, rented or used to build profiles. They are not shared except through links and keys that you create.
No advertising trackers
This website loads no analytics, advertising or third-party scripts, and the app has no telemetry.
No AI run on your data by Glideworth
There is no model in the server, and no hosted AI chat. An assistant you choose can connect through MCP, under a key you control.
No certifications claimed
Glideworth holds no security certification and makes no claim of an independent audit. Tests and a published design are what we can point to.
No end-to-end encryption claim
Even in the design, servers can read your plans. Self-hosting keeps them on a machine you control.
Report a vulnerability.
Please do not open a public issue, discussion or pull request for a security problem. Use GitHub's private vulnerability reporting on the repository's Security tab once the repository is public, or emailsecurity@glideworth.com. Plain email is acceptable. Include what is affected (the self-hosted app, the hosted service, the REST API, the MCP server, the Docker image), the version, the steps to reproduce with synthetic data, and what an attacker gains. Do not send real financial data or live credentials.
| Acknowledge your report | within 3 working days |
|---|---|
| First assessment | within 10 working days |
| Fix or mitigation for critical and high severity | aim for 30 days from confirmation |
| Public advisory | after a fix, normally within 90 days of the report, coordinated with you |
These are targets, not guarantees: the project is run by very few people, and we would rather promise what can be kept. You are credited in the advisory if you wish. There is no bug-bounty programme and no payment is promised. We will not pursue action against good-faith research on your own instance or your own account that stays in scope, and we ask you not to test against other people's data or to degrade the service. Tax or calculation errors are bugs, not vulnerabilities: use the tax-rule correction form on the repository. The full policy isSECURITY.md in the repository, and contact lists every address.
Prefer to keep it all on your own machine?
Self-hosting is free, with every feature.