Your inbox, triaged by an agent 24/7.
A Claude Managed Agent that pulls from Gmail, Slack, Granola, and Notion. Classifies by urgency, drafts responses, and maintains a persistent knowledge base that compounds over time.
Anthropic-hosted. No infrastructure to manage. Runs 24/7. Drafts everything, sends nothing. The agent gets smarter with every run.
Four sources in. Five surfaces out.
The agent connects to your communication and work tools via MCP, runs a classification pipeline on every new item, and routes the output to the right surface based on urgency.
Threads, labels, attachments, drafts
Channels, DMs, threads, @mentions
Transcripts, notes, action items
Databases, pages, wiki, tasks
Slack DM with context, suggested action, source link.
Gmail drafts + Slack replies staged for human review.
Meeting action items with owner + deadline.
Morning briefing: priorities, drafts, wiki-enriched meeting prep.
Persistent Notion knowledge base. People, projects, patterns, decisions.
Five steps. Wiki in the loop.
Ingest + normalize.
Pull new items since last cursor. Normalize into a common schema: source, sender, timestamp, content, thread ID, metadata.
Deduplicate + merge.
Same thread in email and Slack? Merge into one item. Content hash skips anything already processed.
Enrich from wiki.
Query the Assistant Wiki for context before classifying. Person pages surface sender history, open items, and communication preferences. Project pages surface deadlines and risks. Pattern pages flag known behaviors.
Classify + score.
Each item gets an urgency tier (P0-P3) informed by wiki context. Within each tier, items ranked by weighted composite: sender importance, deadline proximity, topic relevance, thread heat, staleness.
Route + update wiki.
P0 gets an immediate alert. P1 gets a queued draft. P2 goes in the digest. P3 is logged. At the end of every run, the agent writes what it learned back to the wiki.
Four urgency tiers. Configurable rules.
Every incoming item is classified against your VIP list, active projects, and channel filters. The tier determines what happens next.
Urgent
Action within 1 hour
VIP sender
Deadline today
Escalation language
Respond
Reply expected today
Direct question
Client-facing
Meeting follow-up
FYI
Awareness only
CC'd on thread
Status updates
Announcements
Low
Batch later
Auto notifications
Marketing
Non-urgent internal
Institutional memory that compounds.
The agent maintains a dedicated Notion database as its persistent knowledge base. Every triage run reads from it and writes back to it. Classification gets more accurate, drafts get more contextual, and digests get more insightful over time.
Person
Communication context for individuals. Preferences, recent interactions, open items.
Organization
Company/team context. Key contacts, active workstreams, relationship health.
Project
Context beyond the PM tool. Decisions, risks, lessons, related threads.
Topic
Recurring themes with evolving positions and cross-team context.
Pattern
Observed behaviors. Email rhythms, escalation cycles, volume spikes.
Decision
Important decisions with context, implications, and status tracking.
Open Question
Unresolved questions with ownership and timeline. Flagged when stale.
Without the wiki, the agent is stateless. "Email from sam@acme.com." With the wiki: "Sam Park at Acme. Prefers email over Slack. Has followed up twice about the API integration. Contract renewal in August."
Maintenance runs on Haiku at 6:45 AM. Staleness detection, confidence decay, orphan detection, contradiction flags. Results feed into the 7 AM morning digest.
The wiki lives in your Notion workspace. You own the data. The agent maintains it. You can read and correct any page.
Your rules. Your priorities.
Five config inputs shape how the agent triages. All editable. All yours.
CEO, founders, board, key clients. Always P0.
Slack channels + Gmail labels to include or exclude.
Notion databases for topic scoring + context enrichment.
Tone, formality, phrases learned from your sent messages.
Sections, order, delivery channel preferences.
What powers it.
System prompt
Detailed instructions that define how the agent triages, classifies, and responds. Customized for your workflow.
6 skills
Email triage, meeting follow-up, daily digest, task extraction, draft response, and wiki maintenance. Sonnet for judgment, Haiku for wiki ops.
Credential vault
Encrypted OAuth tokens for Gmail, Slack, Granola, and Notion. No passwords in code. AES-256 at rest. Automatic token refresh.
Cron scheduler
5 automated triggers. Each fires a session: the agent wakes, runs the right skill on the right model, then sleeps.
Assistant Wiki
A Notion database the agent maintains as its persistent memory. 7 page types: Person, Organization, Project, Topic, Pattern, Decision, Open Question.
Model delegation
Sonnet handles classification, scoring, and drafting. Haiku handles all wiki reads, writes, and maintenance at ~75% lower token cost.
Wakes up on a cron. Sleeps between runs.
Five cadences, each optimized for a different kind of work. Wiki maintenance runs on Haiku before the morning digest. State store remembers where it left off. No items processed twice.
Urgent scan. VIP senders + approaching deadlines. Quick wiki lookup for sender context.
Full triage. All 4 sources, complete pipeline. Wiki enrich + update.
Wiki maintenance on Haiku. Staleness detection, confidence decay, pattern validation.
Morning digest. Wiki-enriched briefing with meeting prep, open questions, draft queue.
EOD wrap. Pending items, stale threads, wiki health, tomorrow preview.
Deploy through Claude Console. Runs 24/7.
Claude Managed Agents
Anthropic-hosted. No infrastructure to manage.
Claude Console
Configure the agent, prompt, skills, and MCP connections. Deploy on your Claude org account.
~$8-20/month
With model delegation. Sonnet for triage, Haiku for wiki. Higher volumes or Opus: $20-35/month.
Drafts, never sends
All responses staged for human review. Nothing sent without your approval.
Build one for your workflow.
The repo has everything: system prompt, skill definitions, MCP setup, cron schedules, wiki schema, cost breakdown, and a step-by-step setup guide. Deploy on your own Claude org account.