some file

General-Purpose Agents

Most AI agents run as software on a desktop or server — Claude Code, Claude Cowork, OpenAI Codex, OpenClaw, and every other coding agent, desktop assistant, or in-house automation your team uses. Install the Netzilo client on the machine, and every one of these agents is monitored across its full activity — file system, processes, MCP tool calls, LLM calls, and network — with no changes to the agent itself.

You don't integrate each agent one by one. The client governs whatever is running on the machine, including agents you didn't build and have no dedicated integration for:

  • Claude Code — Anthropic's command-line coding agent.
  • Claude Cowork — Anthropic's collaborative agent.
  • OpenAI Codex — OpenAI's coding agent.
  • OpenClaw — an open agentic coding tool.
  • …and any other AI agent — desktop assistants, autonomous runtimes, scheduled automations, or anything else making LLM and tool calls on that machine.

How it works

The Netzilo client observes every action an agent takes on the machine — not just its network traffic. File reads and writes, process spawns, MCP tool calls, LLM API requests, network connections, and skill acquisition all flow through Netzilo and are evaluated against the policy you configure. Prompts, tool calls, model responses, and the actions they trigger can be allowed, blocked, redacted, or reported.

Because Netzilo sees the whole sequence — not one request in isolation — it catches the multi-step attacks that single-layer tools miss: prompt injection, credential exfiltration, capability hijacking (an agent loading a skill from an untrusted source, then using what it granted), and data leakage to unsanctioned destinations.

What gets inspected and how it's judged is configured centrally, not per machine:

  • Edge Tools — register and approve the AI tools and MCP servers agents are allowed to reach.
  • Edge Scanners — the rules that inspect traffic and decide what continues, what is rewritten, and what is stopped.
  • Edge Filters — bind tools, scanners, and posture checks into a policy and apply it to a group of machines.

Install on a desktop

Install the Netzilo client, then bring it up:

curl -fsSL https://pkg.netzilo.com/download/linux/install_netzilo_linux.sh | sh
netzilo up

netzilo up opens your browser to sign in. See the full Installation guide for the macOS and Windows installer downloads and platform details.

Install on a server or container

Servers and containers enroll non-interactively with a setup key instead of a browser sign-in — ideal for images, CI, and infrastructure-as-code:

curl -fsSL https://pkg.netzilo.com/download/linux/install_netzilo_linux.sh | sh
netzilo up --setup-key <SETUP_KEY>

Once enrolled, the machine picks up whichever Edge Filter is assigned to its group — so an agent runtime deployed across many containers is governed uniformly, with policy managed in one place.

Browser-based agents

Not every agent runs as a local process — many live in the browser: web AI assistants, in-page agents, and the AI features built into SaaS apps. The Netzilo browser extension brings the same AIDR functionality to browser-based agents that the client brings to local ones — inspecting and governing what they do, with per-domain controls for redaction, download restrictions, prompt-injection blocking, and more.

Installing the client monitors every AI agent on the machine at once, across its whole activity — you don't enroll agents individually. That's the difference from the SDK (one agent, in-process) and plugins (one coding-agent CLI): the client is the broadest, lowest-touch way to bring agents you don't control under full policy.