Munder Difflin

What is Munder Difflin?

Munder Difflin is a local, MIT-licensed multi-agent harness that runs the terminal coding CLIs you already pay for as a coordinated team instead of one session at a time. Each agent is a real terminal process in its own node-pty PTY, and they collaborate through a 'hive' — a local git repo of plain files holding per-agent memory, atomic-file mailboxes, a shared blackboard and an append-only event log. A GOD orchestrator agent (Michael) reads every request, resolves routine ones itself to keep the floor autonomous, and escalates only critical items — spend, destructive operations, scope changes — into an approvals queue. It is aimed at developers whose limit is supervising parallel agents rather than agent capability, and it runs on macOS, Windows or Linux from your own machine.

Key features

• Runs Claude Code, Antigravity (Gemini), OpenAI Codex, xAI Grok, Kimi Code, Gemini CLI, Qwen, OpenCode, Crush, pi.dev, GitHub Copilot CLI, Cursor or a custom command — each in its own node-pty PTY rendered with xterm.js
• GOD orchestrator agent routes tasks and adjudicates traffic, escalating only spend, scope and destructive operations to a human approvals queue; can also be driven by voice
• Hive coordination layer: per-agent memory, atomic-file mailboxes, a shared blackboard, an append-only event log, and a single-committer git design so no agent touches git directly and parallel writes cannot corrupt index.lock
• Optional per-agent git worktrees so parallel agents never collide on branches
• Circuit breaker with a steer to constrain to stop ladder for agents that loop, storm errors or blow their budget, plus per-agent token budgets, real cost computed from transcripts, a durable ledger and OTel spans
• Semantic recall over markdown memory with condensation so the index does not grow forever, plus an enterprise knowledge graph any agent can query
• Command Center with Kanban tasks and dependencies, scheduled missions, live fleet monitoring, a CI watcher, a skills catalog of 227 installable skills, and a built-in Monaco IDE with CHANGES / HISTORY / COMPARE git rails
• BYOK per-provider keys in a write-only secret broker plus local model base URLs for Ollama, LM Studio and vLLM; Slack and webhook entry points can spawn an ephemeral worker, reply in-thread and tear it down

Who it's for

• Getting more parallel throughput out of two or more coding-agent subscriptions you already pay for, instead of running them one terminal tab at a time
• Running a long unattended batch of coding work where the orchestrator handles routine decisions and only pages you for spend, scope or destructive operations
• Supervising several agents on one repository without git collisions, using single-committer routing and optional per-agent worktrees
• Mixing model tiers on purpose — a stronger orchestrator over cheaper workers — while tracking real per-agent cost from transcripts against a token budget
• Running agents against local models via Ollama, LM Studio or vLLM when work cannot leave the machine

When not to use it

If you only ever run one agent session at a time, the entire coordination layer is overhead — the value is fan-out supervision. It is also an Electron app you build and run yourself (Node.js 18+ and a C/C++ toolchain for node-pty's native addon), and the maker's own release notes document real correctness bugs in recent builds — cost reporting that reset its counter on restart and under-reported spend, semantic memory returning NaN vectors on Apple Silicon, and a usage-limit guard in 0.3.8 that never released the agents it held — so pin and update deliberately rather than trusting an old build

FAQ