Every coding agent on your machine runs as you. It can read what you can read, write where you can write, and reach whatever your network reaches. Most of us never decided how much of that it should get. We clicked "allow" a few hundred times, turned on a faster mode to get through a demo, and moved on.
Today we're releasing two free, open-source tools that answer two questions: how risky is my setup right now, and what can a command my agent runs actually touch? They live in one repository, GolemTeam/golem-team-shield, under Apache-2.0. Neither needs an account, and by default neither sends anything anywhere.
brew install golemteam/golem-team/golem-team-shield
golem-check
golem-exec install
golem-check: grade your setup
golem-check reads the settings of the agents on your machine and grades them out of 100, with a score for each area and a list of what to fix, worst first. It covers Claude Code (including managed settings and MDM profiles), Claude Desktop, Codex, Cursor, Gemini CLI, Copilot CLI, Devin, Antigravity, Grok Bot, Hermes, OpenClaw and the VS Code family, plus every repository Claude Code has opened.
What it looks at:
- Run modes. Permission prompts turned off, run-everything modes, YOLO left available.
- Permissions. Allow rules that let an agent run any command, edit any file, fetch any page or call any MCP tool without asking.
- MCP servers. How each server is started and reached: piped from a download into a shell, unpinned packages that fetch new code on every start, plain-http remotes. Tool descriptions are checked where an agent kept a copy.
- Instructions and skills. CLAUDE.md, AGENTS.md, rules files and skills that tell the agent to hide something from you, send secrets out, or ignore its instructions, including text hidden in comments or invisible characters.
- Hooks, secrets and plugin sources. Hooks that run code from the network, credentials written in plain text into agent configs, and plugin marketplaces a stranger controls.
- Protections. Which agents run their commands in a sandbox or a jail.
It reads files and nothing else. It never starts an MCP server to see what it does, because starting a server means running its code. Where no agent cached a server's tool list, it says so by name rather than pretending it checked. If you want those read, -live-tools asks each remote server you've configured for its tool list and checks the descriptions. That flag is the one time it talks to anything, it only sends that request, and it still never starts a local server. Add -pin and the next run flags any description that changed since. It names the file and setting that holds a secret, never the secret.
For teams: -html writes one page to hand to security, -sarif feeds GitHub code scanning, -repo . checks only what a repository brings to everyone who clones it, and the repository doubles as a GitHub Action (uses: GolemTeam/[email protected]) that checks every pull request. -fix shows the change it would make to each risky setting as a diff. -fix -yes makes them, backs up every file first, checks again, and puts everything back if the grade went down.
We ran it on ourselves first
I built this and my own laptop got a C.
Some of that is on purpose. I run my agents in full auto, because we build the tools that make that safe, and golem-check marks any agent that runs commands without asking as dangerous. Two of mine do. Some of it wasn't: it found three places where a leftover allow rule for python3, uv run or bash -c let an agent run any code without asking. I would not have found those by looking.
One thing to know about the grade. A run-everything mode is marked dangerous whether or not a jail is behind it, and a single dangerous finding caps the grade at a C. What stands behind the agent is scored on its own line, Protections: on my laptop, 5 of 7 agents run their shell commands in a jail. If you run agents in auto on purpose, read those two lines together.
Before calling it ready, we also tested the tool itself: fake home directories with 27 planted problems across four agents, a clean setup that should pass, and a dozen real repositories. That found real problems in golem-check. It was quietly skipping projects listed behind folders that no longer existed, so a poisoned CLAUDE.md in the wrong place went unreported. And it rated a Dockerfile style guide as dangerous because the guide explained why curl … | sh is a bad idea. Both are fixed in the release this post points to, with tests that plant the same cases.
It is still pattern rules, not a model. It will miss a clever attack and flag some innocent text. When it gets one wrong, open an issue with the line it fired on.
golem-exec: jail every shell command
golem-exec puts every shell command your agents run inside the operating system's own sandbox: Seatbelt on macOS, Landlock and seccomp on Linux, a restricted token on Windows. golem-exec install sets it up for Claude Code everywhere, and for Codex, Gemini CLI, Cursor and Copilot CLI on macOS and Linux, runs a self-test, and tells you to restart your agents. Your agents still ask you about the command they meant to run, and your allow and deny rules still match it.
With the default settings, a jailed command:
| Can't | Still can |
|---|---|
Read SSH and GPG private keys, .netrc, .git-credentials, shell history | Use the network |
| Read or write browser profiles, chat apps, password managers' data, mail | Use cloud credentials, the SSH agent and Docker (turn these off with the strict profile) |
Write to ~/.ssh, shell startup files, git config, the agents' own settings and hooks | Work in your projects, temp and tool caches |
See secret-shaped environment variables (*TOKEN*, *API_KEY* and others) |
So a prompt injection that gets an agent to cat ~/.ssh/id_ed25519, or to add a line to ~/.zshrc, fails. I had my own agent try the first one while drafting this. The answer was "Operation not permitted."
Three profiles set how tight it is: standard, strict (no cloud credentials) and allowlist, where commands reach only the hosts you list, such as model APIs, package registries and git hosts, through a proxy the kernel enforces. The settings live in ~/.golem/shield.json, which no jailed command can read or change.
What it doesn't cover: it jails shell commands, not an agent's own file-edit or web-fetch tools. And with default settings it stops theft and persistence, not a command sending what it can read to the internet. The allowlist profile closes that.
Why offline, and why read-only
Some scanners in this space start each configured MCP server to read its tools, which runs that server's code on your machine, or send your configs and tool descriptions to a hosted service for analysis. Both are reasonable choices for some teams. We wanted a check you could run on a laptop full of client code without asking anyone first, so golem-check reads files, runs nothing, contacts nothing unless you pass -live-tools, and keeps its findings on your machine unless you pass -upload.
When you want more
The same code runs inside the Golem agent, which is also free to use with no account. golem check gives the same report from the same rules, plus what only a running agent can see. The agent goes further where a single tool can't:
| Open-source tools | Golem agent, no account | Golem with a workspace | |
|---|---|---|---|
| Settings check | golem-check | golem check | Alerts when a risky setting appears on any machine |
| Shell jail | One policy for every command | Locks a session down the moment it reads something untrusted | The same, managed from the console |
| Sandbox for untrusted repos, pages and files | No | Yes | Yes |
| Rules on agents' tool calls | No | Built-in rules | Your own rules and profiles |
| A signed record of what agents did | No | No | Yes |
| What is sent | Nothing by default | Nothing | Your workspace's record, to your workspace |
The docs have the full comparison.