RAHU — your AI coworker, built entirely on Google + Claude Code

RAHU is a Claude Code-native re-implementation of Dimension (the "AI coworker that never sleeps", wound down May 20 2026). It reproduces Dimension's feature set inside the Google ecosystem with no separate UI: every output lands as a Gmail draft/label, a Google Calendar event, a Google Task, or a Google Doc. Claude Code is the engine; Google is the interface.

It learns your writing voice and your tagging conventions from your own email trails, stores them in a local, open-source, token-efficient memory (no SaaS), and routes each job to the cheapest Claude model that can do it well.

---

What it does (Dimension feature → RAHU implementation)

Dimension feature	RAHU	Lands in
Auto email triage — label, prioritize, organize before you open it	inbox-triage agent (Haiku) + email-triage skill	Gmail labels
Auto-draft replies in your tone	voice-drafter agent (Sonnet) + draft-in-voice skill	Gmail drafts
Morning / Evening Briefing (one screen)	rahu-morning-briefing workflow + briefing-writer	Google Doc + draft-to-self
Catch Up — summary + ready draft per message	rahu-catchup workflow	Gmail drafts
Meeting prep & daily recaps	meeting-prep agent + skill	Calendar event / Doc
Action items, all in one place	action-extractor agent (Haiku)	Google Tasks
CRM / deal context + IC memo (VC/founder)	deal-analyst agent (Opus) + deal-context skill	Google Doc + memory
Inbox Autopilot (tag + draft every inbound)	rahu-inbox-autopilot workflow under /loop	Gmail drafts/labels
Learns your voice + tags from history	memory-curator agent + rahu-learn-voice workflow	local memory
30+ integrations via MCP	one open-source Google Workspace MCP server	Gmail/Cal/Tasks/Drive/Docs

Feature list verified by deep research against dimension.dev's own pages (/email, /morning-briefing, /vc, /founder, /pricing), docs, the founder's launch post, and Product Hunt (3-0 adversarial verification on each core claim).

---

Architecture

.claude/
├── settings.example.json     # MCP servers, model routing, hooks, permissions — REVIEW then copy to settings.json
├── commands/                 # what you type
│   ├── briefing.md           # /briefing [morning|evening] [window]
│   ├── catchup.md            # /catchup [window]
│   ├── autopilot.md          # /autopilot [since]   (wrap with /loop for always-on)
│   ├── prep.md               # /prep [today|<meeting>]
│   ├── learn.md              # /learn   (refresh voice + tagging memory)
│   └── memo.md               # /memo <company> [memo|brief]
├── workflows/                # multi-agent orchestration (the /workflows)
│   ├── morning-briefing.js   # triage → draft+tasks+prep (parallel) → synthesize Doc
│   ├── catchup.js            # triage → per-message draft (pipelined)
│   ├── inbox-autopilot.js    # one autopilot pass (loop calls it repeatedly)
│   └── learn-voice.js        # learn voice + tagging from email trails
├── agents/                   # specialists, each pinned to a model
│   ├── inbox-triage.md       (haiku)   ├── action-extractor.md (haiku)
│   ├── voice-drafter.md      (sonnet)  ├── briefing-writer.md   (sonnet)
│   ├── meeting-prep.md       (sonnet)  ├── memory-curator.md    (sonnet)
│   └── deal-analyst.md       (opus)
├── skills/                   # the "how" for each capability
│   ├── email-triage/  draft-in-voice/  morning-briefing/  meeting-prep/
│   └── action-items/  deal-context/    voice-memory/
├── hooks/
│   ├── session-start-inject-memory.sh  # cheap context at session start
│   └── pre-send-guard.sh               # never auto-send unless explicitly opted in
└── memory/                   # local supermemory-style store (the user-context brain)
    └── store.py              # SQLite + optional semantic recall, keyword fallback

How a request flows

/briefing
   └─ Workflow: rahu-morning-briefing
        ├─ phase Triage    → inbox-triage (Haiku)  reads Gmail, applies labels, returns digest
        ├─ phase Act       → parallel:
        │     ├─ voice-drafter (Sonnet) ×N  → Gmail drafts   (recall voice from memory)
        │     ├─ action-extractor (Haiku)   → Google Tasks
        │     └─ meeting-prep (Sonnet)      → Calendar descriptions
        └─ phase Synthesize→ briefing-writer (Sonnet) → one Google Doc

Each agent recalls only the few memory facts it needs (voice, tagging, contact, deal) — never the whole inbox — which is what keeps token cost low.

---

The memory store (local, open-source, token-efficient)

.claude/memory/store.py is a single-file SQLite store inspired by the open-source supermemory project, reduced so it runs anywhere python3 exists — no server, no API key, nothing leaves your machine.

• Namespaced facts (voice, tagging, contact, deal, pref, thread) so each agent pulls only its slice.
• recall --query ... --k N --max-chars C returns a hard-capped compact JSON blob — a recall can never blow your token budget.
• Semantic recall via sentence-transformers if installed; otherwise a deterministic keyword/recency fallback. Same CLI either way.
• Idempotent on (kind,key) — re-learning updates in place, so the profile stays current as your style drifts (the "dynamic" part).

python3 .claude/memory/store.py recall --kind voice --query "investor reply" --k 4
python3 .claude/memory/store.py add    --kind pref  --key no-lp-autosend --text "Never auto-send to LPs."
python3 .claude/memory/store.py stats

Run /learn to populate it from your Sent mail + existing Gmail labels. Optional, for semantic recall: pip install sentence-transformers.

---

Model routing (cheapest model that does the job well)

Job	Model	Why
Triage / labeling / action extraction	Haiku 4.5	high volume, classification-shaped
Voice drafts / briefing prose / meeting prep / voice learning	Sonnet 4.6	writing + synthesis quality
Deal / IC memos, high-stakes drafts	Opus 4.8	deepest reasoning, risk framing

Set in each agent's model: frontmatter and per-phase in the workflows. The session default (settings.example.json) is Sonnet.

---

Setup

1. Install the Google MCP server (open-source, covers Gmail/Calendar/Tasks/Drive/Docs):

   pipx install uv        # or have uvx available
   # create a Google Cloud OAuth client (Desktop), then export:
   export GOOGLE_OAUTH_CLIENT_ID=...   GOOGLE_OAUTH_CLIENT_SECRET=...   USER_GOOGLE_EMAIL=you@domain

   (Zero-setup alternative: the claude.ai-hosted Gmail/Calendar/Drive connectors work without an OAuth app but lack Google Tasks — RAHU then falls back to a Doc checklist.)

2. Activate config: review .claude/settings.example.json, then cp .claude/settings.example.json .claude/settings.json. (Shipped as .example so no permissions/hooks activate without your explicit OK.)

3. Teach it your voice: run /learn once. Re-run monthly to stay current.

4. Use it: /briefing each morning · /catchup anytime · /prep before meetings · /loop --interval 15m /autopilot 20m for always-on · /memo <company> for deals.

---

Safety posture

RAHU drafts, labels, and prepares — it never sends mail or creates calendar invites without you. gmail_send_message and calendar_create_event sit in permissions.ask, and pre-send-guard.sh hard-denies sends unless you create .claude/memory/AUTOSEND_ENABLED. Memory stores distilled patterns, never raw bodies, secrets, or OTPs, and is gitignored. This mirrors Dimension's review-and-send default.