In this video I count down five GitHub repos that climbed GitHub Trending this week, all aimed at one problem: your coding agent can read your SSH keys, open your files and send what it finds anywhere, while most of us keep clicking yes all day. For each repo I explain what it is, how it works step by step, and the catches to know about. It's for developers who run Claude Code, Codex or other agents and want to let them work with less babysitting.
The ranking is by stars gained this week on GitHub Trending's per-language pages (Python, Rust, TypeScript), pulled the evening of Oct 30, 2026 (PT). None of the five shows up on the all-languages page.
What I cover in this video
- 0:00 Why your agent having access to your SSH keys is the real problem
- 0:46 #5 Langfuse: see what your agent actually did
- 2:17 #4 sandbox-runtime: fence one agent in at the OS level
- 4:34 #3 Skill Recorder: teach an agent by doing the task once
- 5:19 #2 OpenShell: a whole fleet of agents under one policy
- 6:50 #1 SkillSpector: is this skill safe to install?
- 9:07 How the five fit together and which one I'd set up first
Seeing and fencing: Langfuse and sandbox-runtime
#5 Langfuse (+509 stars this week)
Langfuse is an open-source platform for tracing and evaluating AI agents. The problem: you hand Claude Code a task, walk away, and come back to a pull request without knowing which files it read or which commands it ran. Here's how the Claude Code integration works:
- Install their plugin (two commands).
- It uses Claude Code's Stop hook, which fires after every response.
- The hook reads the session transcript and turns each turn into a trace, with every tool call, its inputs, outputs and tokens.
- You score those traces with an LLM as a judge, your own code, or by hand.
It's opt-in per project. The catches: content is cut at 20,000 characters by default, errors are mostly silent (check the log file if traces stop appearing), and a developer can turn the hook off in their own config. Treat it as a flight recorder, not a security guard.
#4 sandbox-runtime (+1,093 stars this week)
Anthropic's sandbox-runtime (srt) is the open-source package Claude Code's own sandbox is built on. It locks down files and network for any process at the OS level, without a container. Put srt in front of a command and that command and everything it starts runs inside the fence.
- The fence: Seatbelt on macOS, bubblewrap on Linux, with the network removed completely.
- The proxy: the only way out is a local proxy with an allow list of domains, which starts empty. Anything else gets "connection blocked by network allow list."
- Files: writes are blocked except in folders you allow, and you can block reads on places like your SSH folder.
- MCP servers: put
srt in front of them in your MCP config to sandbox them too.
Anthropic reports sandboxing cut permission prompts by 84% in its own internal use. That matters to me because I run a bunch of agents side by side, and the more you run at once, the less you can approve every command. The catches: it's a research preview so the config could change, Windows support is alpha, and inside Claude Code the sandbox only covers shell commands. Claude's file tools, your hooks and your MCP servers run outside it, so wrap your MCP servers in srt yourself.
Teaching and scaling: Skill Recorder and OpenShell
#3 Skill Recorder (+1,320 stars this week)
Microsoft's Skill Recorder turns a task you do once into a skill your agent can repeat:
- Hit record and do the task. It captures clicks, apps, pages, commands in its own terminal, and your narration, transcribed on your machine.
- GitHub Copilot works out what you were trying to do and the steps in order.
- You review and fix anything it got wrong.
- It writes a
SKILL.md (the same kind of skill file Claude Code reads) or an automation that runs on a schedule.
It prefers real tools like the GitHub CLI over replaying clicks, so one recorded form submission can teach your agent to submit all of them. The catches: it's built for Microsoft's own agents, needs GitHub Copilot access, and hitting Analyze sends your recording to GitHub's cloud, so keep secrets out of it. In their words, the review step gives visibility, not a security guarantee.
#2 OpenShell (+2,352 stars this week)
NVIDIA's OpenShell is a runtime for fleets of autonomous agents. If sandbox-runtime fences in one agent, this runs many behind one set of rules:
- Write a policy for what each agent can touch: files, network and processes.
- Each agent runs in its own sandbox, with the kernel limiting file access and system calls.
- Every network connection is checked against the policy before it leaves.
- The agent never sees your real API keys. OpenShell adds the credential only to requests going to approved endpoints.
- Risky policy changes, like a new host getting your credentials, are flagged with formal verification and held for a human.
Starting Claude Code inside it is one command (OpenShell's sandbox create, then claude), and there's an option for Codex. The catches: you need Docker, Podman or virtualization, the Claude Code quick start needs an API key rather than your subscription login, and it sends anonymous usage data by default (one setting turns it off).
Checking before you install: SkillSpector
NVIDIA's SkillSpector (+2,468 stars this week) answers one question: is this skill safe to install? NVIDIA's own research looked at 31,132 skills and found about 1 in 4 had a vulnerability, with 5% showing likely malicious intent. A skill is just a folder your agent trusts.
- Point it at a GitHub repo, a zip, a folder or a single
SKILL.md.
- Static scan: pattern matching, code structure, data flow tracking, YARA signature rules, MCP tool poisoning checks, and known dependency bugs.
- Optional: a model reads the skill to judge its real intent.
- Risk score: findings add points, critical ones most, and skills with scripts get a multiplier.
- Output as JSON or SARIF with an exit code, so CI can block a risky skill.
It never runs the skill it scans, and you can add it to Claude Code as an MCP server so your agent checks skills before installing them. The catches: NVIDIA calls it defense in depth, not a sandbox, so a clean scan doesn't make a skill safe. If you use the model step, the skill's contents go to that model, so use the no-LLM flag to keep files on your machine.
Key takeaways
- Before install: scan skills with SkillSpector, including any skill Skill Recorder writes for you.
- While running: fence agents with sandbox-runtime, or OpenShell for a fleet.
- After: Langfuse shows exactly what the agent did.
- Set up first: turn on Claude Code's sandbox today (one slash command, built on sandbox-runtime), then add SkillSpector before your next skill install.
- Claude Code's sandbox covers shell commands only, so wrap MCP servers in
srt yourself.
If you're juggling Claude Code, Codex and Cursor across terminal windows, Beckon puts every agent on one canvas side by side. Press a key and talk to them, and each idea gets its own workspace with memory that carries over.
Full video transcript
0:00Your coding agent can read your SSH keys. It can open your file and it can send whatever it finds to any website on the internet. And most of us just keep clicking yes, yes, yes all day long. So this week, five repos climbed GitHub trending that fix exactly that. One lets you see what your agent actually did step by step. One fences it in at the operating system level, and it's the same thing Claude Code's own sandbox is built on. One lets you teach your agent a task just by doing it once. One runs a whole fleet of agents that never even see your real API keys. And number one scans a skill before you install it, which matters more than you think. In this video, we're counting them down by the stars they gain this week. I'm
0:46breaking down how each one actually works and which one I'd set up first. Let's jump right into it. Number five is Langfuse. about 500 new stars this week and almost 37,000 total. So what is it? It's an open- source platform for tracing and evaluating AI agents. In their words, agent evals and observability. Here's the problem it fixes. You hand Cloud Code a task, you walk away, and you come back to a finished pull request. But what did it actually do in there? Which files did it read? Which commands did it run? And where did it go off the rails? Here's how it works with Cloud Code. Step one, you install their plugin, two commands. Step two, it uses Claude Code's stop hook, which fires after every single
1:32response. Step three, the hook reads the session transcript and turns each turn into a trace. Every tool call sits inside it with its inputs, its outputs, and the tokens it used. So you can literally scroll back through a run and see the exact command where things went sideways. And then step four, eval. You can score those traces with an LLM as a judge with your own code or by hand. And it's opt-in per project. Nothing gets traced unless you turn it on in that project's settings. Here's something to be wary about. So, by default, it cuts content off at 20,000 characters. Errors are mostly silent. So, if traces stop showing up, check the log file. And Langfuse says it themselves. A developer can turn a hook off in their own config.
2:17So treat it as a flight recorder, not a security guard. Still juggling cloud code, codecs, and cursor in five terminal windows. Beckin puts every coding agent on one canvas, side by side. Press a key and just talk to them. Every idea gets its own workspace with memory that carries over. Becken. Hey, Beckin.ai. Number four is Anthropic Sandbox runtime. Almost,00 stars this week and almost 7,000 total. And here's why it matters. This is the open- source package that Claude Code's own sandbox is built on. So, what is it? A tool that locks down the files and the network for any process at the operating system level without a container. You put SRT in front of a command and that command and everything it starts runs inside the
3:04fence. Here's how it actually works. Step one, the fence. On a Mac, it uses seat belt, the sandbox that's built into Mac OS. On Linux, it uses bubble wrap and it removes the network completely. Step two, the only way out is through a proxy running on your machine. So, every request walks up to a guard with a list of allowed domains. And by default, that list is empty. Step three, the files. Writes are blocked everywhere except the folders you allow. And you can block reads on things like your SSH folder. So if your agent tries to send something to a site that isn't on the list, it gets connection blocked by network allow list. And then step four, it doesn't just wrap commands. You can put your MCP servers inside it too by putting SRT in
3:49front of them in your MCP config. Anthropic says that in their own internal use, sandboxing cut permission prompts by 84%. That's their number, but it makes sense because once the fence is real, you don't have to approve every single command. And this is the part I care about. I run a bunch of agents side by side in Becken, and the more agents you run at once, the less you can babysit every command, so the fence has to be real. Here's something to be wary about. It's a research preview, so the config could change. Windows support is alpha, and inside Cloud Code, the sandbox only covers shell commands. Claude's own file tools, your hooks, and your MCP servers run outside it. So wrap your MCP servers in SRT yourself. Number
4:34three is Skill Recorder from Microsoft. 1,300 stars this week and over 5 12,000 total. And the idea is so simple. Record yourself doing a task once and it turns that into a skill your agent can repeat. Here's how it works. Step one, you hit record and just do the task. A little bar sits on top of your screen and it captures your clicks, the apps you switch between, the pages you visit, commands in its own terminal, and your voice if you narrate it. Your narration is transcribed right on your computer. Step two, it sends that to GitHub Copilot, which works out what you were trying to do, plus the steps in order. Step three, you review it and fix anything it got wrong. And then step four, it writes a skill, a skill.md
5:19file. That's the same kind of skill file Claude Code reads. or it makes an automation that runs on a schedule. And here's the smart part. It doesn't just replay your clicks. It prefers real tools like the GitHub CLI. So, you can record one form submission and teach your agent to submit all of them. Here's something to be wary about. So, it's built for Microsoft's own agents, and you need GitHub Copilot access to use it. When you hit analyze, your recording goes to GitHub's cloud. So, keep your secrets out of the recording. And in their own words, the review step gives you visibility, not a security guarantee. Which brings us right to number one. But first, number two. Number two is OpenShell from Nvidia. Over 2300 stars this week and almost
6:0421,000 total. They call it the safe private runtime for fleets of autonomous AI agents. So, if number four fences in one agent on your laptop, this one runs a whole fleet of them behind one set of rules. Here's how it actually works. Step one, you write a policy. What each agent is allowed to touch, files, network, and processes. Step two, every agent runs in its own sandbox, and the kernel limits which files it can open and which system calls it can make. Step three, every network connection gets checked against the policy before it leaves the sandbox. Step four, and this is my favorite part, the agent never sees your real API keys. OpenShell adds the credential only to requests going to an approved endpoint. So there's no real
6:50key sitting inside the sandbox to leak. And then step five, when a policy change would open up something risky like a new host that gets your credentials, they use formal verification to flag it and hold it for a human. And starting claude code inside it is one command. OpenShell sandbox create then claude. And there's an option for codecs, too. Here's something to be wary about. You need Docker, Podman, or virtualization. For Cloud Code, their quick start says it needs an API key, not your subscription login, and it sends anonymous usage data by default, which you can switch off with one setting. And number one is Skillspect, also from Nvidia. Almost 2500 stars this week and over 23,000 total. It asks one question. Is this
7:35skill safe to install? And here's why that question matters. Nvidia's own research looked at over 31,000 skills. About 1 in4 had a vulnerability and 5% showed likely malicious intent. That's their research, not mine. But think about it. A skill is just a folder you drop into your agent and your agent trusts it. Here's how it actually works. Step one, you point it at a skill, a GitHub repo, a zip, a folder, or one skill.md file. Step two, the static scan, pattern matching, reading the code structure, tracking where untrusted data flows, signature rules like antivirus, checks for poisoned MCP tools, and known bugs in its dependencies. Step three is optional. A model reads the skill to
8:21judge what it's really trying to do. Step four, the score. Every finding adds points. Critical ones add the most, and skills with scripts get a multiplier because scripts are where the risk is. A low score says safe. A high one says do not install. And then step five, it outputs JSON or SIF and it fails with an error code. So you can drop it into CI and block a risky skill before it ever hits your repo. And it never runs the skill it's scanning. You can even add it to Claude code as an MCP server so your agent can check a skill before it installs one. Here's something to be wary about. Nvidia says it straight up. This is defense in depth, not a sandbox. A clean scan doesn't make a skill safe. And if you use the model step, the
9:07skills contents go to that model. So use the no LLM flag if you want the skills files to stay on your machine. So here's how these fit together. Before you install a skill, scan it with Skillspector. When your agent runs, fence it in with sandbox runtime or open shell if you're running a whole fleet. After it runs, Langfuse shows you exactly what it did. And if you teach your agent with skill recorder, scan what it writes, too. So, which one would I set up first? Turn on clog code sandbox today. It's one slash command and it's built on number four. Then, skillspector before your next skill install. And if you run a bunch of agents side by side like I do, that's exactly what Becken is for at heybecken.ai. So, now I want to hear from you. What's the scariest thing your agent has done
9:53without asking? Tell me in the comments. And if this helped you lock down your setup, a like on this one goes a long way. Have a blessed day.