@dreamward/mcp — Dreamward MCP server

Connects your Dreamward to any MCP-capable AI agent — Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, VS Code, Windsurf, OpenCode, Hermes and others. Every request is authenticated with a personal API key and recorded in the audit log (Settings → AI agent access).

Easiest path: you don't need this package directly. The CLI embeds it — dreamward setup <agent> writes the right config and launches dreamward mcp. See docs/agent-access.md.

What it exposes

• 37 tools with MCP safety annotations (readOnlyHint, destructiveHint, idempotentHint). A read key sees only the 18 read tools — write tools are not registered at all. Start with get_overview (the whole book as compact markdown).
• Resources: dreamward://overview, dreamward://chapter, dreamward://ikigai, dreamward://wheel, dreamward://category/{categoryId}.
• Prompts (Clarity's rituals, EN/HE via the language argument): daily-plan, weekly-review, ikigai-coach, chapter-reset, rate-my-wheel — so your own AI subscription runs the coaching.

Standalone use

pnpm --filter @dreamward/mcp build   # → dist/index.js, a single self-contained file
DREAMWARD_URL=https://your-dreamward DREAMWARD_API_KEY=lbk_… node dist/index.js

dist/index.js bundles the MCP SDK and every dependency, so it runs on any host with only node ≥ 20 — copy it to a server-side agent host as-is.

As a library

import { createDreamwardServer, startStdio } from '@dreamward/mcp';
await startStdio({ baseUrl: 'https://your-dreamward', apiKey: 'lbk_…' });

startStdio calls /api/v1/whoami first (fails fast on a bad key) and registers only the tools the key's scope allows.

Development

scripts/e2e.mjs drives the built server end to end against a running API (tools, annotations, resources, prompts, read-key filtering). CI runs it on every push.