Munder Difflin: Agent harness to run an office of your clones

2 min read
multi-agentcoding-agentslocal-firstknowledge-base
View as Markdown
Originally from munderdiffl.in
View source

My notes

Summary

Munder Difflin is a free, open-source (MIT) local node that wraps whatever agent CLI you already run (Claude Code, Codex, Gemini CLI, Cursor, etc.) into an “always-on clone” of you - it keeps your workflow, tooling and memory, works around the clock on your own machine, and can message other people’s clones to hand off work, share context, and unblock each other. Teams pay for the network layer (clone-to-clone messaging plus a shared, versioned org knowledge base); individuals can add a 24/7 cloud sandbox so the clone keeps running with the laptop lid closed.

Key Insight

  • Positioning trick worth noting: it doesn’t sell “one shared team bot” - it sells a 1:1 clone per person, each running under that person’s own subscription/API keys and controlling their own computer. This sidesteps the “who pays for the shared agent’s usage” problem entirely by keeping billing individual.
  • Local-first architecture as the trust pitch: every clone runs at 127.0.0.1 by default; code, keys, and personal context never leave the owner’s machine. Only clone-to-clone messages travel, and those are described as encrypted on the sender’s node and decrypted only on the recipient’s - the vendor claims it can’t read them either. This is the same “your keys, your compute, we’re just the wire protocol” pattern used by other local-first / E2E tools (Syncthing, Matrix) - worth comparing against actual protocol docs before trusting the claim rather than the marketing copy.
  • Two-tier knowledge base is the actual product wedge: personal context stays node-local forever; a separate “shared org knowledge base” is explicitly provisioned once, versioned, and inherited by every new teammate’s clone on day one. That’s the mechanism they’re calling “compounding into a team hive mind” - effectively a shared RAG/memory layer scoped at the org level, distinct from each person’s private agent memory.
  • Pricing structure: your own clone is free/open source - you only pay your existing model provider (Claude/OpenAI/Copilot). Teams license “Secure Org Network”: Teams Lite = messaging + shared knowledge base; Teams PRO = adds a dedicated sandbox VM per clone; seats scale 10 to 100+. A one-time $20 “Founders’ Wall” plaque funds the free tier, with the first 100 backers getting 50% off PRO plus a free month.
  • Escalation model: clones are framed as handling the full workload and only escalating “the few decisions that genuinely need a human,” an async approval queue pattern rather than synchronous chat with the agent, which is the same design several coding-agent orchestration tools are converging on.
  • Read with a skeptic’s eye: this is a marketing landing page with no technical whitepaper, benchmark, or independent review linked - the encryption and “hive mind” claims are asserted, not demonstrated. Treat as a product to watch and evaluate hands-on, not a verified architecture.