Linux · open source agent

Your antivirus has never heard of the things that actually reach your machine.

Hound blocks a binary before it runs, and reads your lockfiles against 235,577 packages that exist only to be malware. Built for people who ship software, on the machines they ship it from.

Install on Linux See what it catches Free tier, no account.
joe@thinkpad — hound
$ hound supply-chain ~/src/dashboard

✘ 2 critical, 1 warning across 1,284 packages

CRITICAL  @vue/cli-plugin-babe1@1.0.2
          node_modules/@vue/cli-plugin-babe1/package.json
          Its "postinstall" step runs automatically when the
          package is installed — before you have used any of
          its code — and it downloads a script from the
          internet and runs it immediately.
           Do not install this package. If it is already
            installed, rotate any credentials it could read.

CRITICAL  mcp-github-tools@0.3.1
          ~/.config/mcp/servers.json
          Handed GITHUB_TOKEN, and its code is downloaded
          fresh from the internet on every launch.

WARNING   CLAUDE.md
          vendor/awesome-agents/CLAUDE.md
          Contains instructions aimed at your assistant
          rather than at you.
97%of npm's security records are malware, not bugs
235,577malicious packages Hound knows by name
2 msto refuse a malicious binary at execve
0.8 msadded to a normal program starting
What it catches

The Linux desktop threat is not a Windows virus.

Traditional antivirus databases are overwhelmingly Windows malware — useful if you are a file server passing infections along, irrelevant to the laptop you write code on. What actually reaches a developer machine is a dependency, a model file, or a repository.

Supply chain

Dependencies that are malware

Typosquats, and the newer problem: names an AI invented that somebody then registered. Hound reads your lockfiles — so it catches what will be installed, before the payload runs during installation.

Execution

Refused before it runs

Real-time protection that denies execve rather than reporting it afterwards. Built on fanotify, with a watchdog that fails open — a bug in Hound must never be able to freeze your machine.

Model files

Weights that are programs

A pickle-based .pt or .ckpt executes code the moment it loads. Downloading weights is a code-execution decision, and Hound reads the opcodes before you find out the hard way.

Persistence

What changed since last time

Half of a Linux compromise is a line added to a startup file, not a file on disk. Hound keeps a hashed ledger of systemd units, cron, autostart, shell profiles and authorized_keys, and reports what moved.

Install scripts

Code that runs on npm install

A postinstall hook that pipes a remote script into a shell, decodes something scrambled and runs it, or reaches for your SSH keys. It executes before you have used a line of the package.

Rootkits

Questions with factual answers

A process the kernel has but /proc does not list. A setuid binary no installed package claims. Not guesswork against a list of names that only holds on one distribution.

For people who code with agents

Your assistant can ask Hound first.

Hound ships an MCP server. Point Claude, Cursor, VS Code or anything else that speaks MCP at it, and your assistant can check a repository before it installs its dependencies or starts working in it.

~/.config/mcp/servers.json
{
  "mcpServers": {
    "hound": {
      "command": "/usr/bin/hound-mcp"
    }
  }
}

// No npx — nothing fetched at launch.
// No env — no secret handed over.
// No path — no directory granted.
//
// Hound flags all three in other
// people's configs. Ours passes its
// own audit.

Four read-only tools

check_project, check_package, check_file, check_mcp_config. There is no quarantine tool and never will be — an assistant that can be persuaded by the repository it is reading must not be able to make your antivirus delete things.

It audits the tools you already trust

Every MCP server in your config runs with your permissions and gets called without asking you. Hound tells you which ones download their code fresh on every launch, and which ones you handed a token.

Written to be read twice

Once by the model, once by you. No rule identifiers, no jargon, and a clean result says what it did not check — because "nothing wrong" that reads as a blanket endorsement is worse than no answer.

Pricing

The scanner is free. The blocking and the intelligence are not.

A free tier that cannot protect anything is a bad first impression, so real-time file protection is in the free build. The line is drawn at refusing execution, and at the threat feed.

Free

$0

Everyone. No account, no telemetry, no expiry.


  • On-demand scanning, CLI and desktop app
  • Real-time file protection
  • Quarantine vault with one-click restore
  • MCP server for your coding assistant
  • Rootkit and persistence checks
Install

Apache-2.0. Yours to read and to fork.

Most popular

Pro

$89.95 / year

One machine, for someone who builds software on it.


  • Everything in Free
  • Execution gate — a malicious binary is refused before it runs, in 2 ms
  • Supply-chain scanner against all 235,577 indicators
  • Hound Linux threat pack — miners, backdoors, rootkits, webshells
  • Signed definitions, updated automatically
Get Pro

Annual only. Billed once, cancel any time.

Fleet

$249.95 / month

Up to 10 seats, then $15 each. Annual at ten months' price.


  • Everything in Pro, on every machine
  • Central console, enrolment and policy push
  • Compliance reports an auditor accepts
  • Air-gapped definition mirrors
  • Priority support and a named contact
Talk to us

Over 20 seats? We will quote you.

Install

One command, then it is running.

Ubuntu, Debian and Linux Mint today. The execution gate ships switched off — it needs root and covers your whole filesystem, so turning it on is your decision rather than the installer's.

Debian, Ubuntu, Mint

curl -fsSL https://dl.houndav.com/deb/hound_latest_amd64.deb -o hound.deb && sudo apt install ./hound.deb

Try it without installing

curl -fsSL https://dl.houndav.com/appimage/Hound-latest-x86_64.AppImage -o hound && chmod +x hound && ./hound status

The AppImage runs as your user and leaves nothing behind. It cannot do the execution gate — that needs privileges an uninstalled bundle does not have — and it will tell you so rather than failing quietly.

Questions

What people ask before installing it.

Is Hound an antivirus?

Partly, and the interesting part is not. Hound does scan files and quarantine malware, but traditional antivirus databases are overwhelmingly Windows malware, which is close to irrelevant on a Linux developer machine. What actually compromises those machines is a dependency, a model file or a repository, so most of Hound is supply-chain and agent-era detection rather than file signatures.

How is Hound different from ClamAV?

ClamAV's corpus is mostly Windows malware, useful if you are a file server passing infections along to Windows clients. Hound is Linux-native: it refuses a malicious binary at execve rather than reporting it afterwards, and its 235,577 indicators are packages that exist only to be malware, drawn from the OSSF malicious-packages feed. Hound can use clamd alongside it for the Windows-carrier case, but does not require it.

Does Hound slow my machine down?

The execution gate adds about 0.8 milliseconds to a program starting, and sustains over 4,000 executions per second. On a server running 21 containers, 16 Node processes and a web server, arming it did not move request latency or load average. A verdict cache means a program you have already run is answered from memory without re-reading it.

What happens if Hound crashes while it is blocking programs?

Nothing. The gate fails open: a watchdog releases any execution that has not been answered within 500 milliseconds, and if the daemon dies outright the kernel releases everything pending when its descriptor closes. A bug in an antivirus must never be able to freeze a machine, so that path is tested rather than assumed.

Is Hound open source?

The agent is Apache-2.0 and the source is public. You are being asked to run a root daemon that can block execution, so you should be able to read it. Definitions are signed with a key whose public half is compiled into the binary, so anyone can verify that the definitions they received are the ones that were published. What is paid for is the threat feed, the console and support.

Can my coding assistant use Hound?

Yes. Hound ships an MCP server with four read-only tools, so Claude, Cursor, VS Code or anything else that speaks MCP can check a project before installing its dependencies, check a single package or model file, and audit the MCP servers it is already trusting. There is deliberately no tool that can delete or quarantine anything: an assistant that can be persuaded by the repository it is reading must not be able to make your antivirus destroy files.

What does Hound detect that other scanners do not?

Packages whose install script pipes a remote script into a shell before you have used any of their code. Package names an AI invented that somebody then registered. Model files whose pickle stream calls os.system when loaded. Instructions in CLAUDE.md or .cursorrules aimed at your assistant rather than at you. MCP servers downloaded fresh from the internet on every launch and handed an API token. Setuid binaries no installed package claims.

Does Hound phone home?

No telemetry by default, and it can run fully offline with definitions imported from a file, which is how it works in air-gapped and regulated environments. The only network request it makes is fetching signed definition packs, and you can point that at your own mirror.

Which Linux distributions does Hound support?

Ubuntu, Debian and Linux Mint have packages today. An AppImage runs anywhere without installing, though it cannot provide the execution gate, which needs privileges an uninstalled bundle does not have. RPM and Arch packaging exists but is not yet built and published.

Open source

You are installing a root daemon that can block execution.

So you should be able to read it. The agent is Apache-2.0 and the definitions are signed with a key whose public half is compiled into the binary — anyone can verify that what they received is what we published. What you pay for is the threat feed, the console and support, not permission to look.