Skip to content

Guides

Hosting, security and data

Epic Context runs as one stack: a web app, a realtime service, PostgreSQL and S3-compatible file storage. Access is checked on every request, secrets are stored encrypted, every change is recorded and reversible, and the brain's own model calls go through one door an organisation can point elsewhere.

This page is for the architect or security officer who wants to know how it works before an installation. It describes what the code does today, nothing more. Ask Epic Stories for anything it does not answer.

What runs

PartWhat it does
Web appThe book, the settings, the REST API and the MCP connector, in one Next.js server
Realtime servicePresence and the design canvas (Hocuspocus); live block changes reach open pages from the database
PostgreSQL 17 with pgvectorThe graph: blocks, versions, links, people, organisations
Object storage (S3-compatible)Uploaded files and brand assets

The parts ship as Docker images and run as one Docker Compose stack. Database migrations run when the web app starts; a migration that fails stops the start instead of serving a half-migrated database.

The hosted service keeps daily backups: a snapshot of the whole server, and a dump of the database with a copy of the file store, kept for fourteen days.

Running it on your own server

Epic Stories hosts the brain for you today; running it on a server of your own is planned. The stack is built for it: the same images, and the model the brain uses set once per installation (see below). Talk to us if you need it.

Access

  • Invite-only. Nobody gets an account without an invitation to an organisation, and nobody becomes a member without accepting one.
  • Checked on every request. Every page, API route and MCP tool checks the person's role in the organisation before it reads or writes. Someone outside the organisation gets not found, the same answer a missing page gives.
  • Sign-in: passkeys, one-time e-mail links, or passwords hashed with scrypt, with a lockout after repeated failures. Sessions are stored in the database and last thirty days; an admin can end them. See Team, roles and sign-in.
  • AI clients sign in with OAuth (authorization code with PKCE). Access tokens last thirty days; refresh tokens rotate, and reusing an old one ends the chain. A refresh token is spent only after the new tokens are made, so a failure leaves it usable.
  • API keys are stored as hashes and shown once, carry scopes (read, write, sync) and can be revoked at any time. When someone leaves an organisation or is removed from it, the keys they made there are revoked with it; deleting a project revokes its keys.
  • On record. What owners, admins and the suite's own staff do to people and access (roles, removals, invitations, API keys, sign-in help) is kept in a record the organisation's admins read. See Team, roles and sign-in.

Secrets

  • Keys for outside services and model providers, and the signing secrets of installed tools, are stored encrypted (AES-256-GCM). Nobody can read them back after saving.
  • A person enters a key in Settings. A key never passes through a chat, and no MCP tool takes one.
  • Changing what a connection does (its address, sign-in or actions) drops its key, so a key can never be used for something new without a person entering it again.
  • Checked at every start. The server checks that its own secrets are set and that the encryption secret still reads what is stored, and logs each problem on start. An installation can make a problem stop the start instead.
  • Rotated without signing people out. A server secret is replaced by keeping the old value beside the new one for a while (<NAME>_PREVIOUS), so sessions and AI connections keep working; after a leak, the old value is left out and everything signed with it stops.
  • Kept out of the code. Every change to the code is scanned for files and values that look like secrets before it can be built.

What leaves the brain

  • Your AI tools. When Claude, Claude Code or Cursor reads the brain through the connector, the blocks it reads go into that tool's conversation, at the vendor of that tool, under your agreement with them. The connector serves only what the signed-in person may read.
  • The brain's own model calls. The editors, ingestion, extraction and AI-Chat call a model through one door. By default that is Anthropic's Claude. An organisation can choose another provider that speaks the chat-completions format (for example Mistral, Azure OpenAI or its own AI gateway) per agent, per role or for everything, under Settings › Organisation › Connections. A chosen model is a promise: if it cannot be reached, the call fails rather than falling back to another vendor. An installation of its own sets MODEL_PROVIDER, MODEL_BASE_URL and MODEL_DEFAULT once for every organisation on it.
  • Your own Anthropic key. An organisation can save its own Anthropic key (at /<organisation>/settings/ai, owners and admins). Every model call of that organisation then runs on it; a saved key that cannot be read stops the call instead of falling back to Epic Context's key.
  • Public links. A page leaves the organisation only when someone who may edit gives it a public link; locked and legal pages only by owners and admins. The organisation can switch public links off, give them a lifetime and see every live one. See Sharing one page.
  • Outgoing calls. Webhooks and connectors reach public https addresses only, without following redirects, with bounded answers. Webhooks are signed.
  • Exports. You can take the whole brain at any time, as a ZIP of markdown files with the links, messages and images, or as JSON. See Export.

Deleting and keeping

  • Archiving a block keeps it in the project's archive until someone restores it or an owner or admin erases it.
  • Erasing a block takes every version of it too; nobody can restore it in the app. Deleting a project or an organisation erases its blocks the same way. See Deleting.
  • Backups keep a copy for at most fourteen days. That is the only way back after an erase, and it is a step on the server.
  • Demo requests sent from this site are kept for twelve months, then deleted.

Integrity

  • Every write records who, from where and why, and can be undone from the changes feed. See Writes, changes and locks.
  • A locked block refuses every writer, owners and AI included.
  • Writes are validated against their schema at the boundary.
  • An organisation sets a monthly AI token budget; the editors and the suite's apps stop spending at the ceiling.

Sign-in options that do not exist

There is no SAML or LDAP sign-in. Suite apps and AI tools sign in through Epic Context itself, which acts as their OpenID Connect and OAuth provider.

Privacy

The privacy policy for Epic Context and the rest of the Epic Suite is at /privacy.

Next

Last updated: