some file

AI Agent Lifecycle

AI agents act on their own: they call tools, read untrusted content, move data and reach external services. Netzilo AIDR governs what they do, checking prompts, model answers and tool calls against your policy and stopping prompt injection, credential exfiltration, capability hijacking and data leakage before they complete.

How you connect Netzilo to an agent depends on where that agent is in its lifecycle, how much control you have over its code, and where it runs.

The lifecycle

StageWhat it coversHow Netzilo governs it
DevelopmentAgents your team writes and runs in its own processThe SDK, embedded in the agent
Pre-deployment and run-time on a deviceAgents already built: coding-agent CLIs, desktop AI apps, packaged tools, and AI in the browserCoding-agent hooks, the Netzilo client and the Netzilo browser extension
Run-time anywhereAny agent's model and MCP traffic, wherever it runs: workstations, servers, CI jobs, devices you do not manageThe Netzilo AI Gateway

Policy reaches each of them through Filters under Edge > Filters: Endpoint filters for devices, the SDK and the browser, Gateway filters for the users who call the gateway. All of them report to the same dashboard.

Agent types

Custom Agents: you build them

Agents written by your team. Because you own the code, you embed Netzilo in the process with the SDK (Python or TypeScript): one call at startup, then every framework tool call and LLM prompt and answer is governed in-process. This is the development-stage integration.

→ Custom Agents

Coding Agents: you run them

Third-party coding-agent CLIs you use but do not build: Claude Code, OpenAI Codex CLI, Google Gemini CLI and OpenClaw. The Netzilo client registers a hook with each, so their tool calls and prompts are checked before they run, without changing the agents.

Coding agents can use the hooks and the AI Gateway together: the hooks block tool calls on the device as they happen, and the gateway identifies the user, applies model grants and budgets, and counts usage.

→ Coding Agents

General-Purpose Agents: anything on a managed machine

Desktop AI apps, IDE assistants, in-house automations and other agents on a desktop or server. The Netzilo client intercepts the processes you list in an endpoint filter's Agents, on Windows and macOS, and evaluates their AI traffic. On Linux the client does not steer traffic per application: point the app at the client's proxy, or use the hooks or the SDK.

The Netzilo browser extension brings the same governance to AI in the browser: ChatGPT, claude.ai, Gemini and Perplexity. It applies the endpoint filters whose operating systems include Browser, for the signed-in user's groups, and stands down where the client already governs the browser.

→ General-Purpose Agents · Browser Extension

Through the AI Gateway: any agent, anywhere

Point an agent's OpenAI or Anthropic base URL at the Netzilo AI Gateway and give it the user's Netzilo token. The gateway holds the provider keys, applies the user's gateway filters (models, limits, MCP servers, guardrails) and records usage. This covers agents on servers, in CI and on devices without the Netzilo client.

→ AI Gateway

These approaches compose. A team commonly embeds the SDK in the agents it builds, installs the client and the hooks on the machines its developers use, deploys the browser extension, and routes model traffic through the AI Gateway. Endpoint filters govern the devices and browsers, gateway filters govern the users at the gateway, and all of them report to the same Activity log and dashboard.