arjunkshah12345-hash/openpages

OpenPages
Make room for your next idea.

Live demo · GitHub · SuperCompress

Open-source AI workspace with persistent Spaces, collaborative pages,
and SuperCompress as the core context layer.
Open-source alternative to OpenAI Pages / ChatGPT Space.

Quick start · Why SuperCompress · Architecture · MCP · Docker

Most AI workspaces dump the whole room into the model.

OpenPages does not.

Workspace → Retrieval → SuperCompress → Model

Your Space can grow forever. Your context window — and your bill — do not grow with it.

Persistent context without persistent token costs.

OpenPages is a real product (Spaces, Pages, agents, MCP) and a living demo of SuperCompress: query-aware context compression that keeps answer-critical evidence and drops the rest.

Every chat turn, agent run, and MCP workspace_context call:

1. Retrieve relevant pages, files, and conversation snippets
2. Compress that candidate context with SuperCompress (hosted API or local offline stand-in)
3. Infer with your chosen model — OpenAI, Anthropic, Gemini, OpenRouter, or Ollama
4. Show original vs compressed tokens in the Context inspector

SuperCompress is not a plugin toggle. It is the economic layer that makes large agent workspaces viable.

Get a key: supercompress.dev/dashboard · Docs: docs.supercompress.dev

• Spaces — multi-space workspaces (pages, files, search, settings, agent activity)
• Pages — Tiptap block editor, / commands, autosave, checklists, tables, mermaid
• Inline AI — rewrite / improve / shorten / expand with accept / reject proposals
• Workspace agent — cited answers over the Space
• Agent mode — multi-step goals, activity timeline, page creation
• Context inspector — sources, retrieved tokens, SuperCompress savings, expandable context
• Model agnostic — ChatGPT account (no API key), OpenAI, Anthropic, Gemini, OpenRouter, Groq, Mistral, DeepSeek, Together, Fireworks, xAI, Azure, Ollama, or custom OpenAI-compatible — always behind SuperCompress
• MCP — Cursor / Claude Code / Codex can search, read, write, and pull compressed context
• Version history — user / agent / mcp attribution + restore
• Self-host — SQLite by default, Docker Compose included

git clone https://github.com/arjunkshah12345-hash/openpages.git
cd openpages
cp .env.example .env   # optional — onboarding can store credentials instead
npm install
npm run dev

Open http://localhost:3000 → Start a Space runs the local onboarding wizard:

1. SuperCompress — paste your key from supercompress.dev/dashboard (required for the real compiler; offline stand-in is buried behind an intentional confirm)
2. Bring your own inference — pick how you run models:
   • ChatGPT account — Login with ChatGPT (device code) or import ~/.codex/auth.json. No sk- API key for Plus/Pro
   • OpenAI API — paste a platform sk-… key and choose GPT-5.4 / o-series models
   • Other providers — Anthropic, Gemini, OpenRouter, Groq, Ollama, and more
3. Credentials are saved to ~/.openpages/settings.json (mode 0600)

A Product Launch demo Space is seeded so first-run chat already shows compression in the inspector.

You can still use .env instead of (or in addition to) the wizard. Settings file wins when both are set.

Without a SuperCompress key, OpenPages uses a local query-aware fallback so the pipeline and inspector still work. Without a model connection, answers are demo replies — retrieval + SuperCompress still run.

cp .env.example .env
# set SUPERCOMPRESS_API_KEY + at least one model key
docker compose up --build

Data persists in the openpages-data volume (/data/openpages.db).

Expose a Space to Cursor, Claude Code, Codex, and other MCP clients:

{
  "mcpServers": {
    "openpages": {
      "command": "node",
      "args": ["/absolute/path/to/openpages/mcp/server.mjs"],
      "env": {
        "OPENPAGES_URL": "http://127.0.0.1:3000",
        "OPENPAGES_SPACE_ID": "space_demo"
      }
    }
  }
}

Tools: space_search, page_list, page_get, page_create, page_update, file_get, workspace_context (retrieve + SuperCompress).

HTTP bridge: POST /api/mcp with { "tool": "workspace_context", "arguments": { "spaceId": "...", "query": "..." } }.

See ARCHITECTURE.md for the full map.

src/lib/retrieval/      → find candidates in the Space
src/lib/supercompress/  → compress against the query (API + local)
src/lib/models/         → provider-agnostic inference
src/lib/ai/             → shared generate / demo helpers

Next.js · TypeScript · Tailwind · Tiptap · Drizzle · SQLite · Vercel AI SDK · SuperCompress

MIT — build on it, self-host it, star it if SuperCompress made your agent workspace affordable.