Munder Difflin turns your team into an office of always-on AI clones
Two stories today: a local-first multi-agent harness blows up on GitHub and Hacker News, and a home-grown inference trick hints that long-context prefill might be far cheaper than we assumed.
Munder Difflin, a free open-source app that wraps the coding-agent CLIs you already pay for into persistent "clones" of each teammate, took the #1 spot on GitHub Trending and hit the Hacker News front page on Friday. The pitch is local-first multi-agent work: the harness runs on your own laptop, learns how you work — your repos, tooling and notes — and then spins up agents that review pull requests, triage issues and answer questions in your style while you sleep. Clones message each other over end-to-end encrypted channels, so an agent blocked on a design token at 3 a.m. gets unblocked by a teammate's clone instead of stalling until morning. It supports twelve agent backends out of the box — Claude Code, Codex, Gemini CLI, Copilot, Cursor and more — riding your existing subscriptions rather than selling you a new model bill.
Why it matters: most multi-agent orchestration so far has meant trusting a vendor's cloud with your codebase. Munder Difflin bets the opposite way — nothing leaves your machine unless you opt into its paid cloud sandboxes and team network, and with roughly 3,500 stars within a day of trending, the appetite for that trade is obvious. The open question is whether clone-to-clone handoffs produce real merged work or just busy-looking activity; that's the gap between demo and office. But the direction is clear: the agent harness layer is becoming a product category of its own, sitting between you and the models.
A Reddit experiment suggests long-prompt prefill can run roughly three times faster by building the KV cache in independent chunks and stitching the results together. Redditor maddie-lovelace split 256k-token prompts into separate segments, generated caches for each in isolation, concatenated them, and fed the result straight into decode — skipping the single monolithic pass over the whole prompt. With some overlap between chunks on Ling3-tiny's non-KDA layers, the model kept full needle-in-a-haystack retrieval and could even synthesize across the split parts, hitting about 1,300 tokens per second of prefill at 256k context.
The caveats are real: it's one model, one hobbyist setup, simple retrieval tasks, and nobody knows where the quality trade-off hides. But the idea isn't fringe — the 2024 CacheBlend paper showed cached-chunk fusion preserving accuracy for RAG workloads, and this looks like the same principle pushed to extreme contexts. If the approach survives contact with harder reasoning tasks, cheap long-context prefill stops being a data-center-only luxury.
What to watch: whether llama.cpp/vLLM-class runtimes pick up cache-blending as a supported mode — that's the signal it's more than a party trick.
Would you let a clone of yourself answer your teammates at 3 a.m., or does that cross a trust line? Tell us in the comments.
