Vulnetix
Try Vulnetix Request Demo

Your agents, always on, guarded on every path out

A dedicated Belai machine on Cloudflare Sandbox, built from a reproducible Nix image and wrapped in AI, package and egress firewalls. Clef from Cloudflare and Jev from TypeSafe both come included. Pick your repositories and languages, then run sessions, agents and crews from the console.

Start a sandbox See pricing

Every way out of the machine passes a gate

Model calls, package installs and every other request leave the sandbox through a control you set in the console, and Clef or Jev reads what comes back before the model does.

From checkout to a running host in one launch

  1. Start a sandbox. Name it in the console. The machine is created as soon as checkout completes.
  2. Choose repositories. Public GitHub repositories, or private ones through the Vulnetix GitHub App, by branch.
  3. Pick languages. Toolchains and language servers, plus any nixpkgs package, from the signed cache.
  4. Launch. The image is verified, the firewalls go up, your repositories are cloned and Belai comes online.
  5. Work from the web. Run sessions, agents, crews and schedules from the Belai pages you already use.

Clef and Jev decide every call, both included

Clef from Cloudflare and Jev from TypeSafe are security models built for one job: judging whether the text an agent is about to act on is safe. On a Pix Sandbox one of them is Belai's classifier from the first session. Clef is the default and runs on Cloudflare Workers AI, so no Cloudflare key exists on the machine. Switch a sandbox to Jev from its launch settings. Either one reads prompts, file contents, web pages and tool output before the model sees them, and catches prompt injection, jailbreaks and instructions hidden in repository text. The API key is added at the network edge, for TypeSafe's endpoint alone, so the machine holds only a placeholder.

A dedicated container that stays on

Every Pix Sandbox is its own Cloudflare container with a 20 GB disk, running around the clock, in one of three sizes. Small has 1 vCPU and 4 GiB and runs 1 web session and up to 7 agents. Medium has 2 vCPU and 6 GiB and runs 2 web sessions and up to 14 agents. Large has 4 vCPU and 12 GiB and runs 4 web sessions and up to 28 agents. Change size from billing: a running sandbox waits for its agents to finish, then restarts on the new size. It starts with the internet off, opens only the hosts your policy allows, and is reached through the Vulnetix console alone, with no SSH and no open port.

A reproducible image, built with Nix

The sandbox image is defined by a pinned Nix flake, so the same inputs give the same image, byte for byte. Language packs install as signed closures from the binary cache: Python, Node.js, Go, Rust, Java, Ruby, C and C++, .NET, Zig, Dart and cloud CLIs, plus any nixpkgs package.

Every model call through the AI Firewall

The Vulnetix AI Firewall is switched on at launch. Model calls pass through it, provider keys sync into its vault and stay off the machine, and your policy's blocked patterns, PII redaction, model allowlists and spend limits apply. Belai's permission rules, OS sandbox and classifier stay on for every session.

Malware and exploitable versions stop at the registry

Each language you add is pointed at the Vulnetix Package Firewall during launch. Known malicious packages, typosquats and compromised maintainer releases are refused, and versions with known exploited, weaponised or high EPSS vulnerabilities are blocked by your thresholds, before a single file reaches the disk.

See every host your agents reach, and decide

Each request out of the sandbox is decided by hostname. The console lists every host it tried, how often, and how often it was refused, with Allow and Deny one click away. Log mode records everything; deny mode refuses every host except the ones you allow. Jev and both Vulnetix firewalls stay reachable in either mode.

Verify every layer yourself

Put Belai to work around the clock

Start a sandbox See pricing

Questions

What is a Pix Sandbox?

A dedicated Belai machine that runs around the clock on a Cloudflare Sandbox container. You choose the GitHub repositories it clones and the languages it installs, and you run sessions, agents, crews and schedules from the Belai console.

What sizes are there?

Three, each billed monthly per sandbox. Small, from $85, has 1 vCPU and 4 GiB and runs 1 web session and up to 7 agents. Medium, $180, has 2 vCPU and 6 GiB and runs 2 web sessions and up to 14 agents. Large, $350, has 4 vCPU and 12 GiB and runs 4 web sessions and up to 28 agents. Every size has a 20 GB disk. You pick a size when you buy and can change it from billing.

Which decision models are included?

Both Clef and Jev. Clef, from Cloudflare, is Belai's classifier on every Pix Sandbox by default and runs on Cloudflare Workers AI, so no Cloudflare key exists on the machine. Jev, the security model from TypeSafe, is included too, and a sandbox can switch to it from its launch settings. Jev's API key is added at the network edge for TypeSafe's endpoint alone, so the machine holds only a placeholder.

What protects the machine?

Three firewalls are up before the first session: the Vulnetix AI Firewall for model calls, with provider keys in its vault; the Vulnetix Package Firewall for every installed language, which refuses known malware and exploitable versions; and an egress firewall that decides each request by hostname, in a log mode or a deny mode.

How is the image verified?

The image comes from a pinned Nix flake. Belai and the Vulnetix CLI are fetched by the SHA-256 their releases publish, every package comes from one pinned nixpkgs revision, each image is built twice and ships only when both builds match, and CycloneDX and SPDX SBOMs are published. On every boot the sandbox hashes its binaries against the pins before Belai starts.