Brainstack Self-hostedFor IT teams
Brainstack Self-hosted for IT
What you'll run, how it's locked down, and how to set it up.
Architecture
How the server is laid out.
One Linux server or VM, on-premises or in your own AWS, Azure or Google Cloud account, with Docker Engine 28.0 or later. Everything runs in containers on that server, including the local model, which runs on CPU.
Scroll the diagram sideways to see all of it.
Your IdP, your rules
People sign in through Google Workspace, Microsoft Entra or any OIDC provider you already run. Nobody gets a separate Brainstack password.
Reading stays local
Extraction and verification run on a local model (qwen2.5 7B, on CPU) inside the server, on a network with no route out. Your text is never sent to a model provider.
Encrypted where it rests
Provider access is sealed per person with keys you hold. Backups are encrypted to recipients whose secret keys never sit on the server.
Integrations
Sources, sign-in and assistants.
Connect each source once. People only ever see what they could already open in the original app, and removing someone cuts their access everywhere at the same moment.
Sources read-only sync
- GmailLabels you pick
- Google DriveDocs, Sheets, Slides, PDF, text and CSV, in folders you pick
- OutlookFolders you pick
- OneDriveWord, Excel, PowerPoint, PDF, text and CSV, in folders you pick
AI chatsClaude, ChatGPT and Microsoft 365 Copilot, turned on by your admin. Only the person who had a chat sees its facts. Needs Claude Enterprise (Compliance API), ChatGPT Enterprise or Edu (Compliance API), or Microsoft 365 Copilot licences (Microsoft Graph)
Sign-in your identity provider
- Google WorkspaceInternal app in your org
- Microsoft EntraSingle-tenant app
- Any OIDC providerSuch as Okta, Auth0, Keycloak or Authentik
Assistants over MCP
- ClaudeSearch and save team memory
- ChatGPTSearch and save team memory
- Microsoft 365 CopilotSearch and save team memory
- Any MCP clientStandard OAuth, per-person access
- Local modelReads your sources on the box
Assistants connect with standard OAuth and each person's own access, and look facts up by calling search_knowledge. Claude supports MCP. ChatGPT and Microsoft 365 Copilot support MCP connectors in some plans and settings, so check what your plan allows before you roll it out.
Specs
The details, in one place.
- Runs on
- One Linux server or VM, on-premises or in your own AWS, Azure or Google Cloud account, with Docker Engine 28.0 or later. Everything runs in containers on that server.
- Local model
- qwen2.5 7B, running on CPU in its own container. No GPU is needed, and there is no GPU configuration.
- Sizing
- By default the AI service is sized at 4 CPU cores and 12 GB of memory. The database and app need more on top of that. These are starting points to size for your load, not measured minimums.
- Sources
- Gmail, Google Drive, Outlook and OneDrive. Read-only sync of the labels and folders you pick. AI chats: Claude, ChatGPT and Microsoft 365 Copilot conversations, through each vendor's enterprise export API, turned on per vendor by an admin; facts from a chat are visible only to the person who had it. Requires Claude Enterprise (Compliance API), ChatGPT Enterprise or Edu (Compliance API), or Microsoft 365 Copilot licences (Microsoft Graph). Files: Google Docs, Sheets and Slides; Word, Excel and PowerPoint; PDF; and text/plain, text/markdown and text/csv files, in Drive and OneDrive.
- Sign-in
- Google Workspace (internal app in your org), Microsoft Entra (single-tenant app), or any OIDC provider such as Okta, Auth0, Keycloak or Authentik.
- Assistants
- Claude, ChatGPT, Microsoft 365 Copilot or any MCP client, with standard OAuth and per-person access. They look facts up with
search_knowledge. - Network
- The edge serves TLS on your domain and only published routes. Outbound calls go through an egress gateway with an exact allow-list.
- Offboarding
- Sessions, assistant tokens and provider access end in the same step. After restoring an old backup, one command re-applies removals before anyone can sign in.
- Audit
- Every assistant connection, removal and role change writes its audit record in the same transaction as the change.
- Health
manage.py doctorchecks the image, the database roles, the network isolation, sign-in and the backup keys, and names exactly what to fix.- Backups
- Encrypted backups stream straight into GnuPG. The runbook walks you through restoring onto a fresh machine.
- Upgrades
- Migrations run in a one-shot step, one transaction each, before the app starts. If anything looks off, it stops and tells you how to go back.
- Logs
- The database logs only the name of a failed rule, never the row. Customer text stays out of the logs.
- Access
- People only see what they could already open in the original app. Removing someone cuts their access everywhere at the same moment.
Setup
Four things, then it's yours.
A guided setup wizard walks you through it, and we give you the exact settings for Google or Microsoft. After that, your team just signs in.
- A Linux server or VMOn-premises, or in your own AWS, Azure or Google Cloud account, with Docker Engine 28.0 or later. The local model runs on CPU, so no GPU is needed. Start with 4 CPU cores and 12 GB of memory for the AI service, plus room for the database and app.
- Your domainPoint a name like brain.yourcompany.com at it. Certificates are automatic.
- A sign-in appCreate an internal app in Google Workspace or Microsoft Entra, or use any OIDC provider. We give you the exact settings.
- Run the wizardClaim the site, invite your team, connect mail and files.
Tested like someone is attacking it
We tried to break it. Repeatedly.
Before every release, independent reviewers who didn't write the code try to leak data across teams, forge sources, stall the server and misread the docs. Each finding is fixed and re-checked before anything ships.
Questions about your setup?
Get in touch and we'll set up a short call. If it's a fit, we'll help you set Brainstack up.