Hacker News new | ask | show | jobs
Show HN: Kuse Cowork – An open source, BYOK alternative to Claude Cowork (github.com)
2 points by rctstudio2018 158 days ago
Claude Cowork is a fascinating idea for non-devs, but I wanted a version that was open-source, lightweight, and strictly model-agnostic.

So I spent the last 24 hours with Claude Code and here it comes:

Native Rust, Zero-Wrappers: This isn't just another agent wrapper. It’s a ground-up implementation in Rust. No heavy dependencies, no Python bloat, and no reliance on OpenCode/AgentSDK. Just raw performance and a tiny binary.

Docker Sandboxing: Since agents execute code, security is paramount. Open Cowork runs commands inside a transient Docker container.

BYOK (Bring Your Own Key): Use OpenAI, Anthropic, or run entirely offline with Ollama/Local LLMs. You own the keys and the privacy.

Skills, and MCPs: It can already handle complex document tasks (PDF, Excel, etc.) out of the box.

The Development Story: The most interesting part is that I had zero Rust experience before this weekend.

It was a surreal experience: an AI agent (Claude Code) helping me build a faster, secure, and open-source version of "itself."

The project is live on GitHub:https://github.com/kuse-ai/kuse_cowork

It's still very early and experimental, any feedback is welcome :)

2 comments

Cool. Another AI project written in Rust! The local model has been a great idea for privacy and compliance! Which local models do you think can support this kind of agent? I suppose you need at least a 140B model to work well?
Yes, you'd need some kind of edge computing to actually cowork with for this architecture. For now, I would recommend stick with some Tier 1 models like Claude sonnet 4.5, GPT 5.2, etc.

But I highly believe open source models will catch up soon and there are lots of potentials there comparing to closed cloud models.

Love the security-first approach with Docker sandboxing - that's often an afterthought in agent wrappers. The BYOK model is compelling too for teams with compliance requirements.

Curious: how are you thinking about policy enforcement for what the agent can actually do within the sandbox? Like limiting which MCP tools it can call or what parameters are valid?

We're working on governance/guardrails for MCP pipelines at keypost.ai - would be interested to explore if there's complementary overlap. Congrats on shipping!

Thanks!

That’s a good callout. So far we are replying on the model to reliably follow the MCP definitions and scopes.