@Teknium I found cloud sandboxes or an isolated box as the safest options
English
Roey Ben Chaim
153 posts
Roey Ben Chaim
@roeybc
building & securing agents | Staff Engineer @zenitysec | ex msft | climbing, surfing & eating hummus
nyc Katılım Ağustos 2009
492 Takip Edilen117 Takipçiler
Do you agree? Where do you do most of your dev work?
I personally have a linux server machine in my house that I run hermes-agent on and do my dev work from.
How bout you?
Ryan Carson@ryancarson
100% of dev is going to be done in sandboxes in the cloud, controlled by kanban boards. Trust me, I love my local machine and gorgeous mac apps, but all of it is just a terrible form factor for running a team of agents effectively.
English
if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt
Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!
English
@interplato @arvidkahl it's on research mode, you need to be an admin, have a teams/enterprise edition, and not have zero data retention enabled.
I don't know about desktop but if you don't wanna handle github you can use claude code on the web and use this - anthropic.com/news/claude-co…
English
@roeybc @arvidkahl I can’t even find how to do this in Claude code desktop - any tips?
English
What do you use to have Claude Code do a full security sweep of your codebase?
My go-to is “Run a deep OWASP security sweep of the full app, all APIs and any internal services. Report in descending severity and suggest solutions using the AskUserQuestion tool.” in plan mode.
English
This is a great advice not just for benchmarks, but generally. Simon Willison called it Hoard things you know how to do simonwillison.net/guides/agentic…
dex@dexhorthy
We get stuck every month or two on complex problems, usually around complex concurrency problems across multiple services. After solving it manually I always stash the git sha (before/after) in a running list. We now have some very useful eval sets for when new models come out. Most of them still unsolvable without hindsight steering from many many context windows worth of investigation and reproduction
English
@galnagli So cool! All with claude code with custom skills and mcps?
English
For general software tasks? Probably
In skillsbench arxiv.org/abs/2602.12670 we saw software was already pretty saturated, and guessing it’ll be the same for other sectors soon.
It is a great mechanism for non public data that models are not trained on. From interval apis to undisclosed company ip.
English
I give it about 3-6 months before any kind of skills.md file is also pointless.
The same thing happened to vector databases and langchain and every other 'product' built in the narrowing gap of model competencies.
@levelsio@levelsio
Thank god MCP is dead Just as useless of an idea as LLMs.txt was It's all dumb abstractions that AI doesn't need because AI's are as smart as humans so they can just use what was already there which is APIs
English
@arvidkahl Just giving my take on the tech! 🤠
I’m definitely not trying to gatekeep - it might actually win on developer experience and adoption
English
This is the Hacker News usage of "just".
And that is the enemy of all usability, adoption, and ease of use for folks beyond the most technical.
We're entering a phase of hypercompetition, both in what software is written and how it is made.
Let's cool it with the gatekeeping.
Roey Ben Chaim@roeybc
@fatih You can just use the api with info from the swagger or wrap it with a cli. mcp means that every service provider has to add abd maintain another layer on top of that with debatable benefits (context rot)
English
I love the MCP vs CLI/API argument. Some people are completely brainwashed either way. It's like Mac vs PC from 20 years ago, or iPhone vs Android.
In the end who cares, just use what you want. One is better, one is worse.
No amount of debating will change people's opinions
Rhys@RhysSullivan
MCP sucking is a harness problem, not an MCP problem MCP unlocks behavior that is fundamentally impossible to get via CLI or APIs Bad auth, too much context usage, all get solved with an execution layer - your agent writes code to progressively discover and call tools
English
@trq212 We’re still early! I think it’s because a lot of the problems skills solve are manageable (poorly though) in other ways. It’s more under the hood kinda thing.
English
@valhalla_dev I mean, if the VM sits a host you care about. There were vulnerabilities in the past that allowed for vm escape, and also depends on how you do it can have access to mounted volumes.
English
@roeybc in what world is an attack surface smaller on another box versus running it on QEMU on a local machine
English
I'm begging people to just learn how to run QEMU VMs
Perplexity@perplexity_ai
Announcing Personal Computer. Personal Computer is an always on, local merge with Perplexity Computer that works for you 24/7. It's personal, secure, and works across your files, apps, and sessions through a continuously running Mac mini.
English
@thisritchie To be fair I’m talking about the merits of using mcp. If it’s here to stay as a standard or not is a separate question.
Funny analogy though 🤠
English
