Hey from Tobias,
Two things shipped this week that give agents something they did not have before: their own machine, and their own money. Both are open source.
Three topics below, plus a new video on why local AI is worth the trouble.
Your Agent Should Never See Your SSH Keys
Most people run their coding agent straight on their laptop. That agent can read your SSH keys, your other repos, your browser profile, and every credential sitting in your home folder. It's usually fine. The problem is you have no way of knowing the one time it's not.
agentsbox solves this the boring way. Each agent runs in its own rootless container that gets destroyed after every session. It sees your project folder and the mounts you explicitly allow. Nothing else. Credentials load as read-only secrets that never show up in logs.
It works with Claude Code, Codex, OpenCode, and Pi, and it is free. Built by a teammate at one of my clients. If you run agents on a machine that has anything valuable on it, this is the cheapest security upgrade available to you right now.
We Gave Agents a Bitcoin Wallet
An agent that can pay for things needs somewhere to keep the money. That is the piece almost nobody has built yet.
I have been working with JAN3 on agentic AQUA, released open source. It is an MCP server and a command line tool that lets an agent hold and move assets on Bitcoin and Liquid. Point Claude Code at it and your agent has a wallet it can actually use.
Honest framing: Do not put your savings behind it yet. The direction is what matters. An agent that can hold and spend value on its own is a different kind of tool than one that can only ask you to approve a payment.
DeepSeek's Cheap Model Beat Its Own Flagship
DeepSeek-V4-Flash-High now sits at #7 overall in the Frontend Code Arena with 1,586 points, and third among open models. That is a 154 point jump over the preview build, and 121 points above DeepSeek's own V4-Pro-Preview.
Read that second number again. Their fast, cheap model now scores higher than their previous Pro model.
By category it lands #4 in Consumer Product, #6 in Reference-based Design, Data and Analytics, and Gaming, and #7 in Brand and Marketing. The official API is live in public beta, speaks the Responses API format natively, and is already adapted for Codex. Setup is in their docs.
Same pattern I keep pointing at. The gap between open and closed models keeps closing, and it closes from the cheap end first.
The Real Advantage of Local AI (It's Not What You Think)
Everyone argues about local AI on speed and cost, and on both of those the cloud usually wins. That is not why I run models on my own hardware. This video goes through the reasoning process with practical examples.
Watch on YouTube →Sandbox your agents before you fund them.
Tobias
