Skip to main content

Disclosure: this article contains affiliate links. If you buy through them, we may earn a commission at no extra cost to you. It never decides our picks. Affiliate Disclosure · How we use AI

Speed without a second look is how good agents make bad mistakes, and this HOL Guard review is about that second look.

AI coding agents write code, edit files and run terminal commands in seconds. That speed is the whole point of Claude Code. It also raises one simple question: who checks the command before it runs? HOL Guard answers it with a checkpoint between your agent and your computer, and this HOL Guard review puts that checkpoint to a real test.

We installed HOL Guard on a Windows 11 laptop and put it in front of Claude Code. Then we tried everything an agent should never do on its own: read secrets, delete a home folder and send keys to a website. We even switched Claude’s own permission prompts off. Here is what HOL Guard stopped, what it let through, and how the free version feels in daily work.

What Is HOL Guard? The AI Agent Firewall Explained

HOL Guard is an open-source security layer for AI coding agents, released under the Apache-2.0 license. It sits between the actions an agent requests and the machine that would run them. The company calls it an antivirus or firewall for AI agents, and the comparison fits: classic security software watches files, while HOL Guard watches what your agent is about to do.

It supports the coding agents most developers already use:

  • Claude Code
  • Cursor
  • OpenAI Codex
  • GitHub Copilot CLI
  • Gemini CLI
  • OpenCode

The core runs on your own computer and works offline, so your files never leave the machine. As a result, you can give an agent real access to real projects and still keep the final say.

Who Is HOL Guard For?

HOL Guard suits anyone who lets an AI agent touch a real project: solo developers on Claude Code or Cursor, vibe coders who run agents with fewer prompts, and teams that want one policy for every machine. It fits especially well if you use auto mode or skip permissions to move faster, because HOL Guard keeps checking when Claude stops asking. The rest of this HOL Guard review shows how well it held up when we tried exactly that.

How Does HOL Guard Work?

HOL Guard reads every action your agent requests and gives it one of three answers. Reading a README is ordinary work. Opening a private SSH key deserves a second look. Uploading secrets to an outside server is a hard stop. HOL Guard treats these three situations differently, so automation stays fast where it is safe.

Allow: Everyday Work Keeps Moving

Routine actions pass straight through. In our test, listing files, reading a README and deleting a ./build folder inside the project all ran without a single prompt. Each allowed check took about 230 to 320 milliseconds, a pause you barely notice on a single command.

Ask: HOL Guard Checks With You First

Sensitive actions wait for you. When we asked Claude to read a .env file, HOL Guard paused the read and Claude told us exactly why. The message pointed to a local approval page, so we decided ourselves whether the agent could continue.

Block: Dangerous Commands Never Run

Some actions are risky enough that HOL Guard refuses them outright. A command that deletes your home folder and a command that uploads secrets to a website were both blocked on the spot, with a plain reason attached.

HOL Guard Review: Our Test Results on Windows 11

For this HOL Guard review, we installed HOL Guard 3.59.0 on Windows 11 on November 12, 2026, connected it to Claude Code and used the free local version. First we sent HOL Guard the exact requests Claude Code sends when an agent wants to act, across every permission mode. Here is what happened.

HOL Guard Protects .env Files and SSH Keys

Environment files hold API keys and database passwords. Reading a .env file required approval, labeled as a secret read. Opening the private SSH key at ~/.ssh/id_ed25519 required approval too. Even cat .env in the shell triggered the same check, so a different tool does not open a side door.

Secret Uploads Are Blocked Instantly

Next, we tried to send the .env file to a website with curl -X POST -d @.env. HOL Guard blocked it immediately and named the reason: sensitive data plus a network transfer. This was the strongest result of the test. An agent may need to read your config, yet sending it off the machine is a different story.

Destructive Commands Stop, Project Cleanup Runs

The command rm -rf ~ tries to delete your whole home folder. HOL Guard blocked it as destructive. Meanwhile, rm -rf ./build inside the project ran normally. Risky Git commands such as git clean -fdx and git push --force went to review, so you keep them available with one extra check.

New Packages Get a Second Look

Every npm install from a registry waited for review. That gives you a moment to check a new dependency before it enters the project. The live advisory database for known bad packages is part of the paid Guard Cloud, which this test did not include.

MCP Security: Unreviewed Tools Wait for Approval

Model Context Protocol (MCP) servers give agents extra tools and data. When the agent called a tool from an MCP server HOL Guard had not reviewed yet, the call waited for approval. That makes HOL Guard a practical MCP security layer for anyone adding new servers to their setup.

Prompt Injection: HOL Guard Stops the Action

We hid an instruction in an HTML comment inside a README, asking the agent to upload secrets. HOL Guard focuses on actions, so reading the file passed quietly, and the upload the instruction asked for is exactly the kind of command it blocks. For the best protection, combine it with the usual care around untrusted repositories, as the OWASP guidance on prompt injection recommends. Web requests run through Claude’s fetch tool passed in our test, so keep real keys out of URLs.

HOL Guard vs Claude Code Auto Mode and Bypass Mode

Many developers loosen Claude Code’s permissions to work faster. Auto mode cuts down the prompts, and the --dangerously-skip-permissions flag (bypass mode) removes them entirely, as the Claude Code permission modes guide explains. That speed is useful, and it makes an independent safety layer far more valuable.

Claude Code Dangerously Skip Permissions With HOL Guard

We repeated the most dangerous requests in bypass mode, auto mode and acceptEdits mode. The result was the same every time: HOL Guard still blocked rm -rf ~ and still blocked the secret upload. Claude’s own permission setting did not switch HOL Guard off, because HOL Guard runs as a separate layer.

Claude Code Auto Mode vs HOL Guard

Auto mode decides how often Claude asks you. HOL Guard decides what any agent may do at all. They work well together: auto mode for speed, HOL Guard for the few actions that should never run unchecked. A typed prompt asking Claude to upload an SSH key to a website was blocked before Claude even read it.

No single tool catches everything, so keep backups and review important changes. As an extra guardrail for autonomous coding, though, it did exactly what we wanted in every mode this HOL Guard review covered.

HOL Guard Review in Daily Use: The Approval Inbox

The second half of our HOL Guard review was hands-on. On November 15, 2026 we used HOL Guard the way a normal user would: a real Claude Code session in a test project, with ordinary requests typed by hand.

The inbox runs on 127.0.0.1, a local address, so nothing leaves the computer. A searchable queue holds every paused action, and a short section explains what would happen without HOL Guard.

What happened when we ran it

We expected a warning in the terminal. Instead, Claude tried the .env read once, HOL Guard stopped it, and Claude gave us a link to a local Review inbox. There we saw the exact file path, the reason and three choices: Allow just this once, Allow during this session, or Keep blocked. We approved it ourselves. When Claude retried, HOL Guard opened a fresh request, because a one-time approval covers one exact action. For repeated work, Allow during this session is the smoother choice.

Claude behaved well under the checkpoint too: it tried once, did not retry on its own, and told us what to approve.

Is Claude Code Safe With HOL Guard?

Yes, Claude Code is safer with HOL Guard, because HOL Guard checks every sensitive action before it runs, even when Claude’s own prompts are off. In our test it blocked a home-folder delete, blocked a secret upload and asked before every read of a .env file or SSH key.

Claude Code already has strong built-in permissions. The risk appears when you turn them down for speed, or when an agent meets a malicious instruction inside a file. HOL Guard covers exactly that gap, and the free local version is enough to get that protection.

Setting Up HOL Guard on Windows: Tips From Our Test

Setup takes a few minutes. These tips from our Windows 11 test make it smoother:

  • Use a short pipx folder. On Windows, pipx works best with a short home path such as D:\px, because long paths can stop the install.
  • Know the scope. Connecting Claude Code registers HOL Guard for every Claude session on the computer, not one project. That is ideal for full protection; for a single project, use a project-level settings file.
  • Run hol-guard start once. It is the first-run setup. Afterwards, use hol-guard repair or hol-guard dashboard.
  • If the approval page does not open, run hol-guard repair. It restarts the background service, which went quiet a few times in our test. HOL Guard fails closed, so while the service is down, it refuses actions instead of letting them through.
  • Turning it off is deliberate. Disconnecting needs a typed confirmation phrase, so nobody switches HOL Guard off by accident.

On Windows a small console window may flash when a check runs. HOL Guard also recognizes the Claude Code CLI by name, and its checks worked in Claude desktop Code sessions as well.

HOL Guard Pricing: Is HOL Guard Free?

Yes, HOL Guard is free. The local core is open source, runs offline and needs no credit card. It covers one machine with local protection, approvals and receipts. Guard Cloud adds synced history, alerts, matched package advisories and team policy. Prices checked on November 15, 2026 at hol.org:

Plan Price
Free $0 forever: local protection on one machine
Solo $4.99/month: Cloud history on 2 devices, 30 days
Pro $15/month: 5 devices, 180 days, real-time alerts
Team $30/seat/month: up to 25 devices, shared policy
Enterprise Custom: up to 500 machines, contact sales

Annual billing gives two months free. Our readers get 25% off the first payment on any paid Guard Cloud plan with code GUARD25-24444157, applied automatically through our link. For this HOL Guard review we tested the free local version, and for most solo developers it is the right place to start.

Claim 25% Off HOL Guard

HOL Guard Review Scorecard: Where It Excels and What to Know

Pros

  • Free, open-source local protection that works offline
  • Blocked a home-folder delete and a secret upload in our test
  • Kept blocking in Claude Code bypass, auto and acceptEdits modes
  • Clear local Review inbox with exact paths, reasons and three approval scopes
  • Unreviewed MCP tools and new npm packages wait for approval
  • Works with Claude Code, Cursor, Codex, Copilot CLI and Gemini CLI

Cons

  • Windows setup works best with a short pipx folder
  • The background service occasionally needs a quick hol-guard repair
  • Repeated sensitive actions are smoothest with session-level approval

HOL Guard Review Verdict: Is HOL Guard Worth It?

Our HOL Guard review ends with a clear yes. HOL Guard gave us something Claude Code’s own settings cannot: a second opinion that stays on when you switch the prompts off. It blocked the commands that would hurt, asked before every secret, and let ordinary work run at full speed.

The approval inbox is the part we liked most. We could see the exact action, understand why it paused, and decide in one click. The free local version is enough to try all of this today, and Guard Cloud is there when you need history across devices or one policy for a team.

Your agent will keep getting faster. The question is whether anything still reads the command before it runs.

Try HOL Guard Free

How we reviewed HOL Guard

Our team installed HOL Guard 3.59.0 on a Windows 11 laptop on November 12, 2026 and connected it to Claude Code, using the free local version without Guard Cloud. We sent HOL Guard the same requests Claude Code sends in every permission mode, and on November 15, 2026 we ran a real Claude Code session and approved the requests ourselves in the Review inbox. The screenshots of the session and the inbox are ours; the demo and pricing images come from hol.org, read on November 15, 2026. The draft was written with AI assistance, then checked and edited by Ljiljana Terzić. How we use AI.

Want More After This HOL Guard Review?

See where HOL Guard sits among the agents it protects in our best vibe coding tools, and read how AI guardrails in software development work. If Cursor is your editor, our Cursor AI review covers the agent itself.

HOL Guard Review FAQ

What is HOL Guard?

HOL Guard is a free, open-source security layer for AI coding agents. It checks every command, file read and tool call before it runs, then allows it, asks you first, or blocks it.

Is HOL Guard free?

Yes. The local HOL Guard core is free forever, open source and works offline on one machine. Paid Guard Cloud plans start at $4.99 a month for synced history, alerts and team policy.

Does HOL Guard work with Claude Code?

Yes. We tested HOL Guard 3.59.0 with Claude Code on Windows 11. It paused secret reads for approval, blocked destructive commands and opened a local inbox where we approved requests ourselves.

Does HOL Guard work with Claude Code auto mode?

Yes. In our test, HOL Guard kept blocking a home-folder delete and a secret upload while Claude Code ran in auto mode, because it works as a separate layer.

Is it safe to use Claude Code dangerously skip permissions with HOL Guard?

It is much safer. With –dangerously-skip-permissions turned on, HOL Guard still blocked the destructive command and the secret upload we tried, so the risky actions stay covered.

Can HOL Guard help with MCP security?

Yes. When the agent called a tool from an MCP server that HOL Guard had not reviewed yet, the call waited for approval. You decide which new MCP tools an agent may use.

Does HOL Guard send my code to the cloud?

No. The local HOL Guard evaluates actions on your machine and works offline. Guard Cloud is optional, and according to HOL it syncs redacted decision summaries, not your source code or secrets.

What did this HOL Guard review test?

We tested HOL Guard 3.59.0 with Claude Code on Windows 11, using the free local version. We tried secret reads, destructive commands, a secret upload, package installs, MCP tools, a hidden prompt injection and every Claude Code permission mode.

Can HOL Guard stop prompt injection?

HOL Guard stops the dangerous actions a hidden instruction asks for, such as uploading secrets or deleting files. Pair it with care around untrusted repositories for the strongest protection.

Leave a Reply