Oread Studio

A connected home for the story only you can write.

Oread is a writing studio where your world and your words live in one place — and where an AI collaborator who actually knows your book sits beside you as you write. Not a blank box with a chatbot bolted on. A studio that remembers your characters, protects your canon, and never puts a word on the page you didn't choose.

For novelists, screenwriters, poets, essayists, worldbuilders, and roleplayers.

---

---

Why Oread

Most AI writing tools forget everything the moment you close the tab, invent plot you never agreed to, and overwrite your voice with theirs. Oread is built on the opposite instincts:

• It remembers. Every world holds your premise, setting, characters, structure, and the running memory of what you've decided — so the AI answers from your story, not a generic one.
• You stay the author. Nothing reaches your manuscript unless you accept it. The AI proposes; you dispose.
• Your canon is law. Established facts are treated as immutable truth. The AI is told, on every request, not to contradict them.
• Your work is yours. Your world exports as a single clean file you can keep, move, or walk away with. No lock-in.

What you can do

Build a world. Give your story a home: its premise and themes, the setting and its rules, a full cast of characters (with backstories, voices, wants, and wounds), your chapter structure, and a living memory of every decision you've made. It's the series bible you always meant to keep — except this one talks back.

Write, chapter by chapter. A clean, focused editor that dresses itself for the job — novel, short story, screenplay, poetry, essay, or chat/RP — each with its own typography and rhythm. Your prose autosaves as you type, and Save Draft commits it the instant you want the reassurance.

Work with AI in five modes — each with a clear contract about what it's allowed to do:

Mode	What it does
Discuss	Talk it through with character. Nothing gets written — pure thinking partner.
Co-write	Trade turns inside a scene. Insert what lands, ignore what doesn't.
Draft	A full pass from your outline. Review, then insert.
Edit	Rewrites your text as redlines. Accept or reject each one.
Critique	Margin notes and proposed lines. Applies nothing — safe to ask anything.

Chat in character. Talk to anyone in your cast, in their own voice — and they only know what your story says they know. Ask your antagonist why she did it. Interview your narrator. Pressure-test a character before you commit them to the page.

Let the AI do research. In Discuss and Draft, flip on Research and the AI can look up real places, history, and science on the live web to ground a scene or a nonfiction argument — and it cites its sources.

Never lose a version. Before the AI changes a single line of your prose, Oread quietly snapshots what was there. You can always see what came before.

Your words stay yours

• Explicit saves, no surprises. Your world saves when you say Save World. Chats are yours to keep or discard — nothing is stored until you Save Chat.
• Private by design. Each writer's work lives in its own isolated space in the database — no shared tables, no mixing.
• Encrypted keys. If you bring your own AI provider key, it's encrypted at rest and never written into your world file — only a private pointer is.
• Take it with you. Export any world as a single JSON file, anytime.

Bring your own AI (or none)

Oread works with the AI you choose — Anthropic (Claude), OpenAI, Amazon Bedrock, Cloudflare, or a local model via Ollama. Add a provider key once, pick a model for your world, and you're writing. No key yet? The studio still runs end to end so you can explore before you commit.

---

For developers — architecture, setup, and contributing

A connected writing workspace for fiction and nonfiction authors. Build "worlds" (a single JSON document holding premise, setting, entities, structure, and memory), write manuscripts chapter-by-chapter, and work with AI in five modes — including chatting in-character with any character in the cast.

See SPEC.md for the authoritative build spec and its Settled Decisions, and prototype-and-design/ for the design source of truth.

Monorepo layout

packages/shared     TS types + world JSON Schema + format specs (shared by server & web)
apps/server         Fastify + node-postgres — auth, storage, credentials, AI, export
apps/web            Vite + React + TS — the workspace UI (modular, not a monolith)
db/sql              Public schema, provisioner (3-level model), migration functions

Architecture highlights

• Schema-per-user Postgres isolation. public holds only users + sessions; every user gets their own namespace of tables, created by provision_user_schema() inside the signup transaction. Requests run via withUserSchema (sets/resets search_path; validates the schema name).
• Content model: world → named manuscripts (hold format) → chapters (the prose). One world open at a time.
• Save cadence: the world persists only on explicit Save World (+ discrete events); chapter prose is the only debounced-autosave writer (also flushable via Save Draft); chats save only on Save Chat.
• Storage backends: one WorldStore interface, two impls — Postgres (default) and file (OREAD_STORAGE=local). Identical JSON document either way.
• Envelope encryption (AES-256-GCM) for provider credentials + TOTP secrets; master-key versioning for rotation.
• AI: provider adapters for Anthropic, OpenAI, Cloudflare, Bedrock, and local (Ollama), behind one interface; per-mode context assembly with a token budget; server-side mode-contract enforcement; native web search in Discuss/Draft; a deterministic mock fallback when no credential is configured.
• Snapshots are JSON-Patch deltas (+ occasional full). A revision snapshot is taken before any AI-applied change to prose.

Setup

npm install
cp .env.example .env      # fill in DATABASE_URL, MASTER_KEY_V1, SESSION_SECRET

Generate secrets:

openssl rand -base64 32   # MASTER_KEY_V1
openssl rand -base64 48   # SESSION_SECRET

Optional: SIGNUP_WHITELIST=a@x.com,b@y.com restricts who can register (leave unset for open signup).

Database

db:bootstrap / db:migrate are the ONLY commands that run DDL — they are opt-in and never run implicitly.

npm run db:bootstrap      # install public schema + provisioner + migration fns
npm run db:migrate        # apply pending migrations across all user schemas (idempotent, restartable)

The migration runner loops public.users.schema_name, applies each pending idempotent plpgsql migration in its own transaction, bumps schema_version, and is restartable. Run it in Render's pre-deploy command.

Develop

npm run dev:server        # Fastify on :8080
npm run dev:web           # Vite on :5173 (proxies /api → :8080)

Test

npm test                  # all workspaces

DB-integration tests (schema isolation, migration idempotency/restartability) skip when DATABASE_URL is unset. To run them against a scratch database:

DATABASE_URL=postgres://… npm --workspace @oread/server test

Storage modes

• Default: Postgres (schema-per-user).
• OREAD_STORAGE=local: worlds live as ./data/worlds/<id>/world.json — one cat-able file per world, for single-user / offline use. (Auth still uses Postgres.)

Copyright 2025/2026 Claudette Raynor. This software is provided for non-commercial, personal, and educational use only.