Munder Difflin releases open-source harness that wraps Claude Code into "your own clone"

The clone absorbs your workflow, your tools, and the knowledge you've built up over time. That means the longer you run it, the more a newly spun-up clone already understands how you work from the very start. The GitHub repository is published under the chaitanyagiri account, and you can grab the download from the releases page.

Why a harness, and why now

A "harness" isn't the AI model itself — it's the framework that connects that model to an actual work environment. A recent example that captures this trend well: NVIDIA wrapped its own harness around Claude Opus 5 and pushed benchmark scores for that same model from the 30% range up to 100%. In other words, how you wrap a model is becoming just as important as the model itself in determining real-world performance.

Around the same time, AWS released a bridge that lets agents running in the cloud access tools on a user's local computer, and Cursor rolled out 31 official plugins connecting business tools like Gmail and Salesforce to agents. Munder Difflin sits in that same broader trend, but it takes a different tack: instead of building one shared bot for the whole team, it creates an individual clone for each team member.

How the clones actually work

Each clone is a node running on that person's own computer. If one person's clone gets stuck on a task, it can message another teammate's clone to get the information it needs. For example, if a colleague's clone sends a question like "How does the billing service work?" at 3 a.m., your clone can answer on your behalf without waking you up.

Clones review teammates' pull requests according to your own standards and habits, fix bugs, keep an eye on CI, and keep documentation up to date — carrying out their assigned roles around the clock. Only the handful of decisions that genuinely require human judgment get escalated to the user; everything else gets handled through clone-to-clone communication.

How to try it

① Where to start — Download the node file from the releases page of the github.com/chaitanyagiri/munder-difflin repository.

② Step by step — First, pick a CLI agent you're already using — Claude Code, Codex, Cursor, or similar — and connect it, carrying over your existing API key or subscription account. Then run the node to create a clone that learns how you work. If you have teammates, you can set it up so their clones exchange messages with yours.

③ Who can use it — The personal node is free and MIT-licensed, so you can inspect the code directly on GitHub. That said, you'll still need an existing subscription or API key — Claude, OpenAI, Copilot, and the like — to actually run a clone. If you want clones to message each other and share a common knowledge base at the team level, you'll need a Teams Lite or PRO license, which scales from 10 seats up to more than 100.

④ What you can actually do with it — For instance, a designer's clone could cross-check screens against the design system to catch issues and pull out assets, while a product manager's clone could organize issues and keep boards and docs in sync. A sales rep's clone could draft outreach and keep the CRM up to date — essentially, according to the project, almost anything you can do from a command line on a computer can be handed off to a clone.

Pricing structure

Tier	Runs on	Clone-to-clone messaging	Knowledge base
Personal (free, MIT)	User's laptop	Not supported	Personal only
Teams Lite	User's laptop	End-to-end encrypted	Shared team knowledge base
Teams PRO	Dedicated sandbox VM per clone	End-to-end encrypted	Shared team knowledge base
Cloud + Network	Dedicated sandbox VM, runs 24/7	End-to-end encrypted	Hosted organization knowledge base

If you want your clone to keep running even after you close your laptop, you can move it to a dedicated sandbox VM under the Cloud + Network license — and, according to the project, you can always move it back to running locally whenever you want.

Privacy and security design

The project's stance is that for a clone to be trustworthy, users need control over where it runs and who can read the messages it sends. Messages exchanged between clones are encrypted at the sending node and only decrypted at the receiving teammate's node — the developers say not even they can see the contents in between. Users can decide for themselves what knowledge is shared across the whole team versus what stays private context for one person, and once a shared knowledge base is set up, it's version-controlled so newly created clones inherit it automatically.

Editor's take

What makes this project interesting is that it flips the default assumption that's held so far — that a team should share one common bot. Most agents companies have adopted up to now follow a model where a single chatbot is split across many users; even AWS's recently released finance-team agent was a single instance the whole organization logs into. Munder Difflin goes the other way, creating as many clones as there are team members and letting those clones message each other. It's essentially porting a human org chart directly into software.

When you actually try to assemble this harness layer, you keep running into the same sticking point — and it's not model performance. It's deciding whose context gets shared, and how much. If you don't draw a clear line between personal storage and shared team knowledge, a clone could end up leaking sensitive code or customer information to another clone without a teammate ever knowing. How tightly a tool like this locks that down with encryption and permission separation is really what separates the serious ones from the rest.

For small dev teams or startups in Korea, the safer starting point is probably testing the free personal version locally with just one or two people first. Hand off repetitive, clearly-ruled-out work — like PR reviews or doc syncing — to the clones first, and keep humans checking anything involving sensitive information like customer data or contracts until you've moved up to a Teams license. In the coming weeks, it wouldn't be surprising to see companies that already own CLI ecosystems, like Cursor or GitHub Copilot, try to bolt on similar "clone messaging" features of their own.