Show HN: I built an open-source alternative to Claude Cowork

wwayneshng about 5 hours ago 4 commentsRead Article on github.com

HI version is available. Content is displayed in original English for accuracy.

Hey HN,

A few months ago, I tried to automate some of my work with the popular AI agent OpenClaw, and then I quickly realized how difficult it is to get it to work with APIs and third-party services securely, which is essential for a lot of work-related tasks.

Then I realized OpenClaw is more of a personal assistant and it was not designed to get actual work done as a coworker. So I started to build Valmis, an alternative to OpenClaw that works with more than 100 apps and services, with security being the priority.

Valmis addresses the security issue by designing a proxy system: dockerized agent runtime can only request the host machine to make API requests by providing the relevant credential ID. The host then makes the actual request and returns the JSON data to the agent runtime. With this design, you can even turn off the internet access of the agent container while making it work.

Our proxy system now supports 100+ business and productivity integrations, including all Google Workspace apps, Slack, Notion, HubSpot, Salesforce, and Figma.

One of the coolest features of Valmis is the automated workflow. You can automate multi-step workflows using our workflow builder. Each workflow can be triggered by cron, webhooks, app events, and it supports conditions and loops.

I'd be happy to answer any questions in the comment section.

⚡ Community Insights

Discussion Sentiment

83% Positive

Analyzed from 340 words in the discussion.

Discussion (4 Comments)Read Original on HackerNews

wayneshng•about 2 hours ago

Other cool features:

- Agents have cross-session memory: Your agents are able to automatically write memory when you tell them anything worth remembering or when it discovers something that might be useful in the future.

- Browser automation: Agents can operate a headless browser, navigate, fill forms, click, read pages, and take screenshots. Browsers are also managed by the host machine, so agents interact with them using proxy.

- Human in the loop: Whenever there is a critical decision to make, the agent pauses and asks the human for a set of options.

- Team knowledge base and skill system.

Overall, Valmis is designed to run on the cloud and collaborate with humans to get work done (Valmis means "done" and "completed" in Estonian). If you have any questions about the project, please leave a comment, and I'll reply to all questions.

wayneshng•about 2 hours ago

Also here is something fun: Valmis is probably the first AI agent that is able to play real chess with legit moves. We all know LLMs are notoriously terrible at playing chess and always hallucinate moves. So we added a tool to the agent called chess-engine, which basically requires the agent not to rely on text generation to produce moves, but instead to produce each move strictly based on the calculation of a lightweight chess engine built in. And AI can be a great (and sportsmanlike!) chess player.

This is an example I created to show how LLMs can actually do rigorous work. We cannot always trust the output generated (or hallucinated most of the time) by LLMs, but if we add a deterministic tool layer and instruct the model to rely only on the tool's output, we can get more accurate results. In this case, the tool used is a lightweight chess engine.

CupofChineseTea•about 2 hours ago

Looks good, I’ll spin up an instance and give it a try. Btw, does it support multiple users? And the chess engine thing is brilliant, I guess you can extend your logic to other fields, can tools be extended by the end user?

wayneshng•15 minutes ago

It doesn’t current support multiuser, but the system is designed to support multiple users with different roles and permissions. I’m also adding a feature to share credentials between teams, which will be rolled out soon. For the tools call question, yes, this will be a generic mechanism for producing rigorous results from agents. I’m planning to extend it to other fields such as calculation, data analysis and deep research. The basic idea is to give zero trust to any outcome generated through LLMs’ text generation to avoid hallucinations, and only trust results from tool calls.