Exagram

Exagram keeps the documents a team already writes — a doc, a spreadsheet, a deck, a saved conversation — in one place, and lets an AI client read exactly the ones you have shared with it. An Owner can require approval before anything is published, every read is recorded, and what one person can read is decided in one place in the database rather than in each screen.

The words this codebase uses, and what each one means, are in CONTEXT.md. Read that first: they are the same words in the schema, the API and the screens.

Features

• One store for every kind of file. Markdown and text docs, HTML pages, chat transcripts, and uploaded PDF, docx, pptx, xlsx and image files. Text is extracted from uploads so AI tools can read them.
• Revisions and approval. Every change is a new Revision with full history and restore. A Publish Policy decides whether new Revisions go live at once or wait for the Owner to approve them.
• Folders and sharing. Folders with subfolders, shared per person or with a Group as view, collaborate or edit. A share follows the latest Revision or is pinned to one. People can ask for access from a file they can't open.
• Links. Expiring anonymous links to one Revision, with an optional view limit, and "Anyone with the link" access with an optional password.
• AI tools over MCP. Claude, ChatGPT, Cursor and other MCP clients connect through OAuth. They can list, read, search and upload files and manage shares, within the scopes the person approved. Each Connection can be narrowed to chosen folders, and it never reaches more than its person does.
• A Ledger of every read. Owners see who read their files, through which Connection or Link, and which Revision. Admins see the whole install in Audit.
• Working together. Comments on Revisions, an Activity feed, in-app notifications and live updates. Each person has a Personal Workspace and can create or join a team.
• Sign-in. Email and password, Google, and Sign in with Apple.
• Runs on your infrastructure. Docker Compose or Helm, with a bundled or your own Postgres and object store (filesystem, S3-compatible or Google Cloud Storage).

How a read works

People, Links and AI tools all read through the same path. Whether a caller may read an Artifact is decided in one place, in Postgres, and every read is written to the Ledger, including Withheld reads.

flowchart LR
  person(["A person in the app"]) --> api
  link(["A Link, or Anyone with the link"]) --> api
  ai(["An AI tool, over MCP"]) --> api
  api["api"] --> check{"May this caller read it?<br/>Grants, Links, Folder Access, approval"}
  check -->|yes| rev["The Revision they may see"]
  check -->|no| withheld["Withheld: the same answer<br/>as a file that doesn't exist"]
  rev --> ledger[("Ledger")]
  withheld --> ledger

Loading

A Withheld read looks like a missing file, so a caller can't learn that a file exists by asking for it. The one exception is a file shared with Anyone with the link: its reader is asked for the password, or told the link has expired. A Connection never reaches more than the person who approved it, and Folder Access narrows it further.

What is in the box

flowchart LR
  web["apps/web"] -->|"/v1, /oauth"| api["apps/api"]
  web -. "previews, second origin" .-> uc["apps/usercontent"]
  api --> mcp["packages/mcp"]
  api --> core["packages/core"]
  mcp --> core
  worker["apps/worker"] --> core
  worker --> parser["packages/parser"]
  uc --> core
  core --> redactor["packages/redactor"]
  core --> pg[("Postgres")]

Loading

Piece	Where	What it does
API	apps/api	The human API at /v1, the MCP endpoint at /mcp, the OAuth server at /oauth
Worker	apps/worker	Extracts text from uploads, passes each Revision through the Redactor, sweeps expired Links and uploads
Usercontent	apps/usercontent	A second origin that serves HTML bodies and the PDF viewer under a strict CSP
Web	apps/web	The screens
Core	packages/core	The API contract, entities, repositories, and the read path that applies visibility and the Ledger
MCP	packages/mcp	The tools an AI client calls
Parser	packages/parser	Text extraction for docx, pdf, pptx and xlsx
Redactor	packages/redactor	The Detector interface and the parser for placeholders such as [EMAIL]

The worker passes every Revision through the configured Detector. The built-in detector reports no Findings; it records the placeholders an author typed, so a Connection's Revision that carries one waits for the Owner's approval.

Running it locally

You need Git, Node 24 with pnpm 11 (corepack enable, or mise install), OpenSSL, and Docker with Compose. The database scripts run psql and migrate in containers on the host network, so on macOS use OrbStack, or Docker Desktop with host networking turned on (Settings → Resources → Network). Linux works as is.

pnpm install
make local-db-up      # starts Postgres on 127.0.0.1:25432
make local-db-init    # writes .env, creates and migrates the database

OAuth, and with it every AI client, needs a signing key. .env starts with an empty key set, which lets the API start but cannot sign anyone in through OAuth. Put a fresh RSA key in it:

sed -i.bak '/^OIDC_JWKS=/d' .env && rm .env.bak
printf 'OIDC_JWKS=%s\n' "$(openssl genrsa 2048 2>/dev/null | awk '{printf "%s\\n", $0}')" >> .env

Then start the apps:

docker compose up

Open http://localhost:5173 and sign up; the first account is the Admin. The app, the API, the MCP endpoint and the OAuth server are all on that one origin: the web app proxies /v1, /oauth, /mcp and /.well-known to the API so the session cookie and the OAuth interaction cookie stay first-party. HTML bodies are the deliberate exception and are served from http://localhost:3002 instead, sandboxed.

To look at the data with a GUI client, connect to host 127.0.0.1, port 25432, user and password exagram, database exagram.

Connecting an AI client

Sign in, open Connections, and copy the server URL (http://localhost:5173/mcp locally). The client sends you back to Exagram to consent, where you choose what it may do; it then reaches everything you reach. You can revoke it later on the same page, which stops tokens it already holds.

Locally, use a client that connects from your own machine:

• Claude Code: claude mcp add --transport http exagram http://localhost:5173/mcp
• Cursor: add an MCP server with that URL.

Claude.ai, Claude Desktop's custom connectors and ChatGPT connect from their providers' servers, so they can't reach localhost. They need an install with a public address; see Connect AI tools.

Running it on a server

docs/self-host.md covers installing Exagram on your own infrastructure with Docker Compose, Kubernetes or Cloudflare.

Contributing

CONTRIBUTING.md covers the checks to run, the conventions, and how pull requests are reviewed. Everyone taking part follows the Code of Conduct.

License

Copyright (C) 2026 Cleverse.

Exagram is licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only); see LICENSE.

To report a vulnerability, see SECURITY.md.