Security

Stackbridge is normally deployed inside your own infrastructure: a Helm chart onto your Kubernetes or OpenShift cluster, including fully air-gapped installations. In that model your knowledge base, your documents and your search queries never traverse a network we operate, because there is no route by which they could.

We also host it for teams that would rather not run it themselves. That deployment is EU-based, and the same product with the same controls.

The AI is a component you select, not a dependency we impose. Point Stackbridge at a commercial model under your own agreement, or at a model running on your own hardware. In the second case no content leaves your network at any point, including the text of the questions your engineers ask it, which is usually the part a reviewer cares about most.

Chat and embeddings can even point at different machines, so a shared GPU box can serve the large model while the small embedding model sits next to the application.

We never train on customer content. Not our models, not anyone else’s.

Tenant isolation

Every query is scoped to the organisation on the caller’s verified session, never to a value supplied by the browser. The main knowledge table is physically partitioned per organisation in the database, so isolation is structural rather than a filter somebody has to remember to write.

Roles

Administrator, member, and a genuinely read-only viewer. Every action that writes is gated, and administration (credentials, members, exports, connectors) is separated from ordinary use.

Single sign-on

Microsoft Entra ID, Google Workspace and GitHub. Local passwords can be switched off entirely, and sign-in can be pinned to a single provider: Entra or Google today, since GitHub works for sign-in but cannot yet be the enforced one. MFA is applied by your identity provider, under your own policy.

Automation credentials

An AI agent or CI job gets its own token, which can never hold administrator rights, is stored only as a hash, can be given an expiry, and is revoked rather than deleted so the record of what it did survives.

Removing someone

Removing a person or reducing their role takes effect on tokens they already hold, within seconds, rather than whenever those tokens happen to expire.

Two logs, and they answer different questions. The audit trail records what changed: who approved, edited, merged or deleted a piece of knowledge, with a before-and-after. The security log records what happened to the system: sign-ins and failed sign-ins, including attempts against addresses that do not exist, permission denials, credential use and configuration changes.

Both are tamper-evident. Each record commits cryptographically to its own contents and to the record before it, so editing, deleting, reordering or inserting a row breaks the chain at a detectable point. An administrator can verify the whole chain on demand and file the resulting fingerprint somewhere the database operator does not control, which is what turns a log from a claim into evidence.

Separately, Stackbridge records which entries were opened: someone opening an entry in the app, or an AI agent fetching one over MCP, so “what did this departing engineer read” has an answer at offboarding. It is always on rather than enabled per organisation. Search results and chat answers quote entry content without opening the entry and are not covered today; we would rather you knew that now than discovered it during an investigation. The record is treated as what it is, monitoring data about employees, so only administrators can read it and it is deleted automatically on a window you set (30 days unless you raise it). If your works council needs the detail before you deploy, ask us for it.

Credentials at rest

Integration credentials (your Jira token, Slack secrets, model provider keys) are encrypted with AES-256-GCM before they are stored, and decrypted only server-side at the moment they are used. There is a documented key rotation procedure.

Retention

Search-query logs, read-access records and security events each expire on a window you configure, purged automatically. Unbounded retention of what your staff searched for is a liability, not a feature.

Getting it back

One request exports your approved knowledge entries and the full audit trail as structured JSON. There is no export fee and no retrieval process to negotiate. Drafts, rejected entries and the operational logs stay in the database. If you need those in the export too, say so and we will put it in the contract rather than imply it here.

Uploads

Files attached to an entry are served so they cannot execute in the application’s own origin, with per-file and per-organisation size limits, and storage keys generated by the application rather than taken from the uploader.

A Stackbridge install can run with no inbound connectivity from the internet, including Slack. Slack integrations normally require a publicly reachable URL for Slack to call; Stackbridge instead dials out and holds the connection open, so the full Slack integration works with nothing exposed. Jira posts to a webhook, so it needs a route the Jira server itself can reach, which for a self-hosted Jira is an internal one. The Microsoft Teams bot is the exception: Teams only delivers over an inbound HTTPS endpoint, so a deployment that must expose nothing is a deployment without Teams. Everything else (Slack, Confluence, GitHub, Bitbucket, the models) is outbound only.

Outbound traffic honours your egress proxy, including your no-proxy rules, so in-cluster calls to your own model or database are not accidentally routed through a proxy that cannot reach them.

Email lennert@versaminds.live. Please include enough detail to reproduce it. We will acknowledge, keep you informed while we fix it, and credit you if you would like to be credited. We will not pursue anyone who reports a genuine issue to us in good faith and gives us a reasonable chance to fix it before disclosing.

We are a young company, and a page that implies otherwise will not survive your review.

Stackbridge is not ISO 27001 certified and not SOC 2 audited. Those certify an organisation after an accredited external audit, and we have not been through one. We build to support your obligations under frameworks like ISO 27001 and NIS2, which obligate operators such as you rather than software suppliers such as us, and we can walk your team through exactly which controls the product provides and which remain yours.

We keep an internal assessment of our own gaps, in writing, including the ones we have not closed yet. If you are evaluating us seriously, ask for it: lennert@versaminds.live. We would rather show you the list than have you find it.

How we handle personal data is in the privacy policy.