AI
Munder Difflin turns CLI agents into AI clones
Munder Difflin is a free, MIT-licensed harness that runs AI agent clones locally, with optional cloud hosting and encrypted team coordination.
Munder Difflin is offering a free, open-source harness that turns command-line AI agents into persistent “clones” of individual teammates. Munder Difflin says the system runs locally by default, uses the subscriptions and API keys users already have, and can coordinate multiple agents across a team.
The project is aimed at more than software development. Its core idea is to capture a worker’s workflow, tools, context, and memory, then let an agent continue operating on that person’s computer. A developer clone could review pull requests, fix bugs, monitor CI, and update documentation; a designer clone could audit screens and export assets; and a sales clone could prepare outreach, update CRM records, and chase follow-ups.
The local application wraps one of 12 supported CLI agent providers: Claude Code, Codex, Grok, Kimi Code, Antigravity, Qwen, Gemini CLI, OpenCode, Crush, Pi, Copilot, and Cursor. Munder Difflin describes the harness as an orchestration layer rather than a replacement model. The underlying agent still comes from the user’s existing provider subscription or API access.
How the clone system works
The setup has three basic stages:
The project’s interface uses an office-themed simulation in which users can monitor agents as they work. Munder Difflin says the simulation itself is deterministic and does not consume tokens; users can also switch to a simpler full-screen view.
The orchestration model divides work among specialized agents. One example on the site shows an orchestrator routing research to Claude Code, implementation to Codex, review to Claude Code, and Git operations through separate isolated worktrees. The project also shows clones exchanging requests and files: one agent can ask another for design tokens, receive the relevant information, and continue toward an open pull request.
That coordination is designed to happen without a human manually relaying context. A clone can ask a teammate’s clone a question, receive an answer, and continue working while the person is unavailable. Munder Difflin says the system is intended to escalate only decisions that require human judgment rather than interrupting users for every blocked task.
Local execution, encrypted messaging, optional cloud
The default architecture is local-first. Each clone runs as a node on its owner’s computer, with code, keys, and personal context remaining on that machine. The project says the local service runs at 127.0.0.1 and that the only information sent between teammates is end-to-end encrypted clone-to-clone messaging.
Munder Difflin says messages are encrypted on the sender’s node and decrypted only on the recipient’s node. Its security material shows X25519 and AES-256-GCM in the message path. The project also separates personal context from shared organizational context: a team knowledge base can contain common workflows, tools, and decisions, while repositories, notes, and individual working styles remain tied to the owner’s node.
Those are architectural claims made by the project, not an independent security audit. Munder Difflin says its node, protocol, and cryptography are available under the MIT license for inspection on GitHub.
The optional cloud service addresses the main limitation of local execution: the laptop must remain on while a local clone is working. Under the Cloud + Network offering, each clone runs continuously in a dedicated sandbox virtual machine, allowing it to keep operating with the laptop closed. Munder Difflin says customers can switch a clone between local and cloud execution.
For teams, the Network service adds encrypted communication between clones on different teammates' computers and a shared organizational knowledge base. The higher-tier Cloud + Network configuration adds dedicated sandbox VMs for each clone and a hosted knowledge base in the customer’s controlled environment. The site says these environments are not shared infrastructure, but it does not provide technical deployment details or name a cloud provider.
Pricing and availability
The individual version is free and open source. Users supply their own Claude, OpenAI, Copilot, or other supported agent subscriptions or API keys, and the harness consumes those providers' existing hourly or usage limits.
The site lists two paid services rather than charging for the core application:
- Cloud: a dedicated sandbox VM that keeps a clone running when its owner’s laptop is closed.
- Network: encrypted clone-to-clone messaging and a shared organizational knowledge base.
Munder Difflin lists Solo Local as free, while Solo + Cloud, Teams Lite, Teams Pro, and larger organization plans require users to contact the company. Team plans scale from up to 10 seats to groups of 20, 50, 100, and 100+. The project does not publish prices for those plans.
There is also a $20 one-time Founding Supporter option, which provides a name on the project’s Founders' Wall. The site says Munder Difflin will remain free for individuals.
The current release is identified as v0.4.5, with release notes citing accurate cost reporting, semantic memory on Apple Silicon, and more reliable agent messaging. The combination of local execution, support for multiple agent CLIs, and optional encrypted coordination positions Munder Difflin as a control layer for teams already paying for coding and productivity agents. Its unresolved commercial detail is the cost of keeping those clones running in the cloud: the project describes the infrastructure, but not the subscription prices.