Munder Difflin
github.com/chaitanyagiri/munder-difflin- Category
- AI Agents
- Rank
- No. 2054Tools index
Previous survey · No. 2065 ·
- Type
- APP
- Use case
- Agent Building · Coding
- GitHub
- 8.3k stars
- Latest release
- v0.5.5
- Date
About
Local multi-agent harness that turns terminal coding CLIs like Claude Code, Codex, and Gemini CLI into a coordinated team of agents with long-term memory and mailboxes, visualized as avatars working on a shared 2D office floor. Runs on your existing subscription limits, with bring-your-own keys or local LLMs.
What it does
A desktop app that runs several terminal coding agents at once, each as a real process in its own terminal, and makes them work as one team. Each agent gets a memory, an inbox and an outbox. A router passes messages between them through a local git repository of plain files, and one lead agent, called Michael, takes your brief, hands out work and asks you only about spending, destructive operations or scope changes. Every agent shows up as a pixel avatar on a 2D office floor, and you can type into any of its terminals.
Why it's ranked here
It drives the CLIs people already pay for instead of replacing them: Claude Code, Codex, Gemini CLI, Grok, Qwen, OpenCode, Crush, Copilot, Cursor and any custom command, plus local models through Ollama, LM Studio or vLLM. It is MIT licensed, ships builds for macOS, Windows and Linux, and has real operational guards in the code, such as cost caps and a runaway circuit breaker. Against that, the project labels itself pre-release and pushes a paid Pro tier hard in its own pitch.
What's good
Coordination avoids a common failure: agents never touch git themselves, and a single committer writes the shared message store, which prevents lock-file corruption. The main process logs and survives stray crashes instead of taking every running agent down. Secrets for integrations are write-only, so the interface never receives a stored key back. Per-agent token caps, a total spend cap and a breaker that trips on repeated tools, error storms or token velocity give you brakes before a bill runs away.
Tradeoffs
Anonymous product analytics are on by default and must be switched off. Its docs disagree with each other: the readme describes an approvals queue for escalations, while the bridge code says there is no approval queue and human-bound messages go to the lead agent. The version string in the manifest lags the advertised release. Some features, including meeting transcription, sit behind the paid Pro tier. The dependency list includes two public tunnelling libraries, worth checking before you expose webhooks.
How to use it well
Best for a developer who already pays for one or more agent CLIs and wants several sessions running in parallel on one machine with shared memory and visible hand-offs. Brief the lead agent rather than each worker, set token and spend caps first, and turn analytics off if that matters to you. Spawn workers in isolated git worktrees when they share a repository. It is not a hosted or team server; everything runs on your own desktop.
Technical notes+
package.json declares an Electron app built with electron-vite and electron-builder (dist:mac, dist:win, dist:linux), depending on node-pty, the xterm terminal packages, Pixi, better-sqlite3, zustand, posthog-node, localtunnel and tunnelmole; its only test script, test:focused, runs the Node built-in test runner over a test folder. Its version field lags the release README.md advertises. docs/ARCHITECTURE.md splits a terminal plane (a PtyManager streaming per-agent IPC channels) from a hive and event plane covering the on-disk mailboxes, the hook server and the semantic memory wrapper. src/main/index.ts installs uncaughtException and unhandledRejection handlers that log and continue, and enableCodexRemoteForSpawn falls back to a local TUI on any failure and skips win32. src/preload/index.ts exposes a typed window.cth bridge, redacts secretRef to hasSecret, and documents telemetryEnabled as default ON.
Observed
- License
- MIT, per the README license badge
- Language
- TypeScript
- Packaging
- Electron desktop app with a React renderer; signed and notarized macOS builds
- Platform support
- macOS, Windows and Linux
- Declared maturity
- Pre-release, per the README status badge
- Agent runtime
- Each agent runs as a real process in a pseudo-terminal via node-pty
- Supported CLIs
- Claude Code, Codex, Gemini CLI, Antigravity, Grok, Kimi Code, Qwen, OpenCode, Crush, Pi, GitHub Copilot CLI, Cursor, custom commands
- Local models
- Ollama, LM Studio or vLLM
- Telemetry
- Anonymous product analytics on by default, opt-out; posthog-node is a dependency
- Pricing model
- Free open source build plus a paid Pro tier that gates features such as meeting transcription
- Test surface
- One focused test script using the Node built-in test runner
Read from README.md, package.json, docs/ARCHITECTURE.md, src/main/index.ts, src/preload/index.ts.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.