Claude Code Command Center Beats OpenClaw via Agent SDK Layers
Build a multi-agent AI hive mind with voice war room and self-managing memory on existing Claude Code—no new frameworks or API costs—using Agent SDK as bridge for ultimate flexibility over lock-in tools like OpenClaw or Hermes.
Layered Architecture on Claude Code Foundation
Mark Kashef rebuilt his personal AI command center directly on Claude Code using Anthropic's free Agent SDK as the core bridge, avoiding new frameworks like OpenClaw or Hermes Agent. The problem: Weekly hype around agent frameworks creates lock-in, extra costs, and migration headaches when better options emerge. Options considered: Adopt OpenClaw (banned by Anthropic TOS for subscriptions post-April 4th), Hermes, or Anthropic Channels (unreliable disconnections). Decision: Stick to Claude Code subscription ($ existing cost), add ~200 lines for SDK-Telegram bridge in V0, then layer modular features. Why? Layers are removable; foundation upgrades with Anthropic (better models, slash commands, memory). Tradeoffs: Upfront hundreds of hours vs. infinite malleability—no house-moving for new frameworks.
V0 was simple: Telegram interface + SQLite DB + SDK for remote Claude Code terminal. Evolved to multi-UI (Telegram, browser dashboard via Cloudflare tunnel, Slack/Discord options), multi-agent (Main triage, Comms, Ops, Content, Research), subsystems (Gemini memory extraction, Pipecat voice). Auto-launch via macOS launchd services spins agents on boot. Mission control dashboard auto-assigns tasks: LLM (cheap Gemini) picks best agent, e.g., thumbnail gen → Content agent using NanoBanana API.
"The best part of this stack was it promised to be infinitely malleable because outside of these 200 lines of code to enable the bridge between the SDK and your service you can add on whatever you want." – Mark on V0's aha moment, emphasizing modularity over rigid frameworks.
Hive Mind Multi-Agent Delegation
Five specialized agents share a "hive mind"—unified memory of all tasks—for cross-awareness without full multi-terminal chaos. Main agent triages: Delegates 9/10 tasks immediately (e.g., YouTube script → Comms agent), knows competencies, notifies via Telegram. From Telegram: "Hey, can you pull the latest emails?" → Ops agent; Main queries hive mind for updates. War room: Browser localhost + WebSocket + Pipecat orchestrates real-time voice convos (Gemini Live STT/TTS, Deepgram alternatives). Delegates mid-convo: Voice task → SDK subprocess → Telegram sub-agent.
Experimental meeting room (Daily.co + Pika avatars) for video agents—expensive, so lean on Gemini Live + Pipecat frames/envelopes. Three routing rules manage voice flow. Message queue prevents silent failures: Queues concurrent messages (user + cron jobs), processes one-at-a-time per agent.
"I'm Maine, Mark's triage and default agent. I handle general requests and delegate tasks to specialized agents to get things done fast." – Live demo quote from Main agent, showing delegation in action during YouTube audience interaction.
Personalized, Self-Managing Memory System
Memory is "very personal," tailored to Kashef's Obsidian vaults/business. Gemini 1.5 Flash (cheap, huge context) acts as "washing machine": Scans chats, classifies facts/preferences/context → SQLite (or Supabase/Pinecone/Obsidian alternatives). Every 30min: Filters, decays (importance distribution), consolidates. Pinned memories (99 general, 122 insights, fundamentals like name/email) persist across agents/time.
Per-agent Obsidian injection: Comms pulls comms folder MD files via Claude CLI skills at session start. Classifier tags for commit/pin. Exfiltration guard blocks unauthorized responses.
"Gemini behind the scenes is acting like a washing machine... it will be able to see what is a fact, what is a preference, and what's a context." – Mark on dynamic memory curation, rejecting one-size-fits-all YouTube memory hype.
Alternatives rejected: Anthropic Channels (frequent disconnects after 2-3 days). Dashboard shows memory stats: Pinned vs. decaying.
Security, Reliability, and TOS Compliance
Chat ID allowlist: Only whitelisted Telegram IDs interact. Multi-layers: Data/logs guards, boot PIN, exfiltration blocks. Not bulletproof, but stackable (future Cloud Mythos?). TOS: Personal use of SDK with Claude Code ok per Boris (Claude creator)—not commercial, no third-party like OpenClaw (banned). Queue + launchd ensure no wake-up disconnects.
"The last thing I want to worry about is opening up my phone after waking up and seeing that Telegram is not connected." – On rejecting Anthropic Channels for reliability.
Philosophy: Invest in Claude Ecosystem
New frameworks? Clone repo, audit, layer features into Claude Code—no full rewrites. "You're investing in your Claude Code ecosystem... Everything else is a luxury." Free blueprint: Mega-prompt, 8 Power Packs, assessment prompt, 20-page architecture guide (Gumroad).
"If Elephant Agent explodes... you can go and clone that repo... take off the shelf what interests you." – On cherry-picking without lock-in.
Key Takeaways
- Start with Agent SDK + Telegram (~200 lines) for Claude Code bridge; layer UIs/memory/voice modularly.
- Build hive mind via shared task memory for multi-agent awareness; main triage delegates to specialists.
- Use Gemini Flash as memory "washing machine" for classification/decay; inject Obsidian per-agent.
- Pipecat + Gemini Live for cheap voice war room; queue messages to avoid failures.
- Reject frameworks for layered flexibility: Anthropic upgrades foundation for free.
- Secure with chat ID allowlists, PINs, guards; auto-launch via launchd.
- Auto-assign tasks in dashboard: Cheap LLM picks best agent.
- Personalize memory—copy bits from YouTube, but fit your life.
- Free blueprint kit replicates: Prompts, architecture guide.
- TOS-safe for personal: No third-party harnesses.