Golem docs

Start

Free tools

Check your AI agents' settings and jail their shell commands with open-source tools, or run the free Golem agent with no account. What each one does, and what it leaves out.

There are three ways to use Golem, and the first two are free with no account. Each one replaces the one before it rather than adding to it.

Open-source toolsGolem agent, no accountGolem with a workspace
Installgolem-team-shield, then golem-exec installThe installer with no keyThe installer with a key, or golem connect
Shell jail for your agentsOne policy for every command, from ~/.golem/shield.jsonA policy per session: a session locks down after a web fetch, a prompt injection or a jail denialThe same, with its settings in the console
Settings checkgolem-checkgolem checkAlso alerts when a risky setting appears on any machine
Sandbox for untrusted repos, pages and filesYesYes
Rules on agents' tool callsThe built-in rules, and golem lockdownYour own rules and profiles
A record of what agents didSigned, in the console, with alerts and investigations
What is sentNothingNothingYour workspace's record, to your workspace
LicenceApache-2.0, source on GitHubFree to use; the binary is not open sourceYour workspace's terms

The open-source tools#

golem-team-shield holds two programs. Install them with brew install golemteam/golem-team/golem-team-shield, download them from its releases page, or install them with Go. Its README has the commands for each system.

golem-check#

golem-check reads the settings of Claude Code (including managed settings and MDM profiles), Claude Desktop, Cursor, Codex, Gemini CLI, Copilot CLI, Devin, Hermes, OpenClaw and the VS Code family of editors, and of each repository Claude Code has opened. It grades them out of 100, scores each area, and lists what to fix, worst first. The areas are:

  • Run modes: permission prompts off, every command allowed, every folder trusted.
  • Permissions: allow rules for any command, file, page or MCP tool.
  • MCP servers: how each server is started and reached, and what its tool descriptions tell the agent, where an agent kept a copy of them.
  • Instructions and skills: text that tells the agent to hide things, send secrets out, or ignore its instructions.
  • Hooks, secrets written into settings, and plugin sources.
  • Protections: which agents run their commands in a sandbox or the jail.

It reads files and nothing else, never starts an MCP server, and names a secret's file and setting, never the secret. It reads a server's tool descriptions from the copy Cursor, Antigravity, the VS Code family or Codex wrote down the last time it started the server, and names every configured server no agent kept a copy of, so you know which ones it could not check. golem-check -live-tools also asks each remote server in your own configs for its tools, with the headers your config already sends it and nothing else, and reports a tool whose description changed since your agent stored it. Add -pin to keep a hash of what each server said, and the next -live-tools run flags any description that changed since, for every agent. It never starts a local server. The Golem agent reads every server's tools as your agents receive them. golem-check -json lists every finding, including the informational ones. golem-check -repo . checks only what a repository brings to everyone who clones it, -html writes one page to hand a security team, -sarif writes SARIF for GitHub code scanning, and -fail-on suspicious exits 3 so a build can stop on a finding. The repository is also a GitHub Action (uses: GolemTeam/golem-team-shield@v0.3.9) that checks every pull request and uploads the findings to code scanning. With a Golem workspace, the action's upload input (or golem-check -repo . -upload) also sends a repository's findings to the console, beside each machine's agent settings, with a CI token minted in IAM that can do nothing else. It is off unless you turn it on, sends from the default branch only, and never fails the job when it can't send.

golem-check -fix shows, as a diff per file, what it would change in your own settings to undo each dangerous or suspicious finding, and changes nothing. It takes out a setting that stops an agent asking, such as bypass mode or an allow rule for any command, and takes away other users' access to a file holding a credential or to a hook script. It never changes a repository's files, managed policy, or a file it can't parse, and it doesn't move secrets. golem-check -fix -yes makes the changes. It backs each file up to ~/.golem/check-backups first, then checks again. If the grade drops, it puts every file back and exits 4. golem check -fix is the same.

golem-exec#

golem-exec is the same shell jail the Golem agent uses, run on its own. golem-exec install:

  1. Copies it into ~/.golem/bin.
  2. Writes ~/.golem/shield.json.
  3. Points Claude Code at it, the same way the Golem agent does.
  4. On macOS and Linux, adds its hook to Codex, Gemini CLI, Cursor and Copilot CLI where each is installed. The hook rewrites each shell command to run through the jail. Codex asks you once to approve the new hook. Antigravity, Hermes and OpenClaw are jailed through the Golem agent, not golem-exec on its own.
  5. Runs a self-test: a plain command must run, and a read of ~/.golem must be refused.

Then restart your agents. golem-exec status shows each agent's coverage and what commands get, and golem-exec uninstall puts everything back.

With the default settings a jailed command:

  • Can't read private SSH and GPG keys, .netrc, .git-credentials or shell history.
  • Can't read or write browser profiles (cookies, saved logins), chat apps, password managers' data, mail or Messages.
  • Can't write to ~/.ssh, your shell startup files, git config or the agents' settings and hooks.
  • Loses secret-shaped environment variables.
  • Keeps the network, cloud credentials, the SSH agent and Docker.

~/.golem/shield.json starts from a profile, which decides the cloud credentials, writes and network; any of those the file names itself outranks it:

ProfileWhat commands get
standard (a new install)Writes kept to the project, temp and tool caches; cloud credentials and the network open.
strictAs standard, without cloud credentials.
allowlistAs strict, with only the allowed hosts reachable: by default the model APIs, package registries and git hosts.

golem-exec install --profile strict picks one. The console offers the same three as quick choices under Sandbox → Settings → Shell commands. The file sets the same choices the console sets for a workspace:

SettingWhat it does
profilestandard, strict or allowlist.
require_jailEvery agent here must be jailed: a command the jail can't hold is refused, enforce: false no longer turns it off, and golem-exec status and golem-check report an unjailed agent as dangerous. An administrator can set it where a user can't, in /Library/Application Support/Golem/managed-shield.json (macOS), /etc/golem/managed-shield.json (Linux) or C:\Program Files\Golem\managed-shield.json (Windows), whose keys outrank the user's file.
enforcefalse runs commands without a jail.
lockdownEvery command loses the network, credentials and writes outside its project. On Windows, commands don't run at all.
ssh_agent, docker, cloud_credentialsThe three grants.
env_allow, env_denyVariables to keep or remove, by name or as a PREFIX*.
writes, allow_write"workspace" keeps writes to the project, temp, the tool caches and the paths you add.
network, allowed_hosts"allowlist" lets commands reach only the hosts you list, through a proxy; everything else is refused, and a refused host is named in the command's output. A list entry can be a preset: @model-apis, @package-registries (npm, PyPI, the Go module proxy, crates.io, Maven Central, RubyGems) or @git-hosts (GitHub, GitLab).

A command can neither read nor write ~/.golem, so the agent being watched can't change these settings. A name the jail doesn't know, or a file that isn't valid JSON, locks every command down until it is fixed. With the Golem agent installed, the agent decides and this file is not read.

What separates golem-exec on its own from the Golem agent is when the jail tightens. On its own the policy is the same for every command, so a command can still send what it can read to the internet unless you turn lockdown on for everything. The Golem agent locks a session down at the moment it reads something untrusted, and leaves your other sessions alone. The limits of the jail are the same either way.

The Golem agent, with no account#

The installer with no workspace key sets the machine up to run on its own:

curl -fsSL https://app.golem.team/install.sh | sh

It installs the Golem agent as a background service, puts the shell jail in front of Claude Code, prepares the sandbox and turns on the built-in rules. Nothing is sent anywhere. See Install with no account.

On a machine that had golem-exec installed, the agent replaces it in ~/.golem/bin under the same names, so Claude Code's setting doesn't change. golem check is the same command as golem-check: the same report and the same options. To go back, uninstall the agent and run golem-exec install again.

The Golem agent is free to use on any number of machines, personal or commercial, under the licence it ships with.

With a workspace#

golem connect --key glk_... links a machine running the agent to a workspace. The agent then signs what each agent does into the workspace's record and follows the workspace's rules and settings from the console. It starts reporting risky agent settings, and it brings in review and the defender. Nothing needs reinstalling.

Updated 2026-11-09 · This page as Markdown