Claude Code Layers Replace OpenClaw and Hermes Agents

Layered Claude Code Foundation Outlasts Hype Frameworks

Mark Kashef ditched OpenClaw and Hermes Agent for a custom setup on his Claude Code subscription, avoiding new API costs and framework lock-in. The core is Anthropic's free Agent SDK, a 200-line bridge running remote Claude Code sessions via Telegram, Slack, or Discord. This evolved from V0 (Telegram + SQLite, one agent) to a full system with five specialized agents (triage/main, comms, content, ops, research) sharing a "hive mind"—a unified memory of all tasks.

Why this over alternatives? Frameworks like OpenClaw tie you to rigid structures; here, Claude Code is the unchangeable base, with all features as swappable layers. "The short version is that my foundation is Claude Code itself, and everything else is a removable layer on top." (Mark explaining rejection of OpenClaw/Hermes). New tools? Clone their repos into Claude Code, audit, and integrate selectively. Tradeoff: Initial ~200 lines of code to bridge SDK, but infinite malleability—no house-moving when trends shift.

Results: Instant delegation (e.g., main agent assigns YouTube script to comms agent, notifies via Telegram). Handles parallel sessions like multiple terminals, from phone or browser dashboard.

Hive Mind Enables Cross-Agent Awareness and Smart Delegation

Agents don't silo; hive mind syncs completed tasks across all, queryable via natural language (e.g., "What has ops agent done?"). Main agent triages: delegates 9/10 tasks based on agent skills, executes only if specified. Mission control dashboard (Cloudflare tunnel) auto-assigns via cheap Gemini model: input task like "Create thumbnail with Nano Banana API, Claude mascot causing havoc," picks content agent, queues for execution.

Delegation flow: Voice/browser input → queue (prevents silent failures from concurrent messages) → classify message → commit to memory → route to agent via SDK subprocess. Message queue ensures one-at-a-time processing despite cron jobs/scheduled tasks.

"I'm Maine, Mark's triage and default agent. I handle general requests and delegate tasks to specialized agents to get things done fast." (Main agent demo in war room, showing real-time handoff to comms for YouTube script). Tradeoffs: Relies on Claude Code's native skills/slash commands; multi-sessions increase local compute but auto-launch via macOS launchd spins them on boot.

Self-Managing Memory Tailored to Personal Workflow

Memory rejects one-size-fits-all: combines SQLite (lightweight DB), .md files, decaying entries, pinned facts (name/address/email), and Obsidian vault injection per agent (e.g., comms pulls comms folder). Gemini 1.5 Flash (cheap, huge context) scans chats every 30min as a "washing machine": extracts facts/preferences/context, classifies importance, decays low-value (distribution gauge shows persist/fade).

Pinned: 122 insights on preferences. General: 99 memories. Exfiltration guard blocks unauthorized Telegram pings. Alternatives considered: Supabase/Pinecone for scale, pure Obsidian. Why Gemini? Monitors all non-private chats dynamically.

"Memory is very personal... my memory system is designed for the way my life works and the way my business runs." (Mark on customizing beyond YouTube trends). "Gemini behind the scenes is acting like a washing machine." (Describing extraction from conversations). Obsidian integration: CLI skills read relevant folders into Claude.md at session start.

Rejected Anthropic Channels: Frequent disconnects after 2-3 days despite stable MCP server. SDK + custom queue = reliable.

War room: Browser localhost + WebSocket + Pipecat (open-source voice orchestration). Live convo with agents assembling in sidebar; speech-to-speech via Gemini Live/Deepgram/Cartisia (leanest: Gemini Live). Delegates voice tasks to Telegram agents via SDK subcommands.

Experimental: Daily.co for Meet-like rooms with Pika avatars (expensive). Pipecat handles frames/envelopes, three routing rules for voice-to-text/agent handoff.

"YouTube hits wider reach and builds authority faster than individual emails or community engagement... Knocking out the big stuff first sets you up for a better day." (Main agent prioritizing day in Daily.co room). UIs: Telegram (/dashboard), browser mission control (add agents, swap models).

Security Layers and TOS Compliance for Production Use

Not bulletproof, but stacked: Chat ID allowlist (only your Telegram ID pings), PIN on spin-up, logs/data isolation, exfiltration guard. Complies with Anthropic TOS (post-April 4 ban on third-party like OpenClaw for Claude subs; SDK for personal local tools OK per Boris from Anthropic).

Auto-launch: launchd services. Modular: Swap backends/interfaces easily.

Future-Proof Philosophy: Invest in Claude Ecosystem

Avoid framework churn: "As new things continually come out... you don't have to worry about doing the equivalent of moving houses." Build skills/processes in Claude Code; layer features (e.g., port Hermes X). Free blueprint: Mega prompt, 8 Power Packs, assessment prompt, 20-page architecture guide.

"You're investing in your cloud code ecosystem on your computer, focusing on making the best skills, slash commands, and processes possible." (Mark on layering over Anthropic improvements). Hundreds of hours iterated; members "swearing by it."

Key Takeaways

Start with Agent SDK + 200-line bridge to Claude Code; pick Telegram/Slack for interface.
Build hive mind for cross-agent sync: unified task memory queryable by all.
Use Gemini Flash for memory washing: classify/decay every 30min, pin essentials.
Queue messages to avoid concurrent failures; auto-delegate via cheap LM in dashboard.
Layer voice with Pipecat + Gemini Live: WebSocket for live browser war room.
Inject Obsidian per-agent; allowlist chat IDs for security.
Reject frameworks for modularity: Clone/audit open-source features into your Claude base.
Auto-launch agents on boot with launchd; expose dashboard via Cloudflare tunnel.
Customize memory to your workflow—personal > viral YouTube systems.
Verify TOS: Personal SDK use OK, no third-party commercialization.