---
title: "Agent Gateway"
description: "Learn what the Agent Gateway is, what it gives your organization, and where to start."
source: "Guardian Platform"
---
The Agent Gateway accounts for every AI agent that works in your environment. It gives you an inventory of your agents, control over what they are allowed to do, and a record of every request they make to an LLM provider.

The gateway is an optional part of a Guardian Platform deployment. Your administrator enables it during deployment.

## Why the gateway exists

Coding agents such as Claude Code, Codex, and OpenCode do their work by calling LLM providers like Anthropic and OpenAI. Left alone, those calls go straight from the developer's machine to the provider. Your organization has no inventory of which agents are running, no way to control what they do, and no record of what they sent or received.

The Agent Gateway sits between your agents and the providers. Each request passes through it, so each request can be attributed, checked, and recorded.

## The two parts

The Agent Gateway has a client side and a server side.

**Viper** runs on the developer's machine, in CI, or next to a service. It wraps or proxies the agent. Developers start their agent through Viper and then work as usual. Viper captures the agent's configuration and state, signs each outbound request with the agent's key, and sends it to the gateway.

**The gateway** runs as part of your Guardian Platform deployment. It verifies each signed request, applies your policies, forwards the request to the provider, records the exchange, and returns the response to the agent.

## What every request gets

Every request that passes through the gateway is:

- **Tied to an identity.** The request comes from a registered agent with its own agent ID (`urn:uuid:...`) and an owner in your organization.
- **Signature checked.** The gateway rejects requests that are unsigned, tampered with, replayed, or signed by a revoked agent.
- **Policy checked.** Any policy plugins your organization or project has configured run before the request reaches the provider.
- **Recorded.** The request, the response, the decision, and a reference to the agent's state at that moment are kept as an audit record.

## What you get

**Inventory.** The Agent Catalog in Guardian lists every registered agent, who owns it, which projects it works in, and when it was last active. See [Agent Inventory](./agent-inventory.mdx).

**Controls.** You decide what agents may do with policy plugins, including OPA/Rego rules such as model allowlists. Policies run organization-wide, with an optional additional layer per project. You can also revoke an agent at any time. See [Gateway Policies](./gateway-policies-and-plugins.mdx).

**Audit.** Every exchange is recorded. You can open any exchange to see what the agent sent, what it received, and why the gateway allowed or denied it. See [Audit Trail](./gateway-audit-trail.mdx).

**Agent configuration.** Viper captures the agent's settings, skills, permissions, and environment when it starts each session and attaches that record to every request. You can see what an agent was configured with for any past request, and how its configuration changed over time. See [Agent Configuration](./agent-configuration.mdx).

**Verification rules.** Configuration validation rules flag risky agent settings, such as unrestricted shell access or a disabled sandbox, on every request. Rules observe and tag; to block an agent based on its configuration, you write a policy, or require a certification of the configuration you approved. See [Agent Configuration](./agent-configuration.mdx#configuration-validation-rules) and [Agent Certifications](./agent-certifications.mdx).

## Supported agents

The gateway supports agents that talk to the Anthropic API, the OpenAI API, or the ChatGPT backend. Viper has a dedicated command for each of these agents:

| Agent       | Command          |
| ----------- | ---------------- |
| Claude Code | `viper claude`   |
| Codex       | `viper codex`    |
| OpenCode    | `viper opencode` |
| OpenClaw    | `viper openclaw` |

Any other agent or application, such as a LangGraph service, runs behind `viper proxy` and points its provider base URL at Viper. See [Run Agents](./run-agents.mdx), or the [LangGraph Quickstart](./langgraph-quickstart.mdx) and [Deep Agents Quickstart](./deepagents-quickstart.mdx) to run one as a named agent, and the [OpenShell Quickstart](./openshell-quickstart.mdx) to run one in a sandbox.

## Where to start

1. Read [Gateway Basics](./gateway-how-it-works.mdx) for a walkthrough of a single request.
2. Follow [Agent Onboarding](./agent-onboarding.mdx), which also lists the supported versions. In short:
   - [Install Viper](./install-viper.mdx) on the machine where your agent runs.
   - Register the agent with Guardian. See [Register Agents](./register-agents.mdx).
   - [Run your agent through the gateway](./run-agents.mdx), or follow the [LangGraph](./langgraph-quickstart.mdx), [Deep Agents](./deepagents-quickstart.mdx) or [OpenShell](./openshell-quickstart.mdx) quickstart, and adjust Viper with [Viper Configuration](./viper-configuration.mdx) if you need to.
3. Check what is and is not recorded for your agent on [Agent Coverage](./agent-coverage.mdx).
4. Find the agent in the [Agent Catalog](./agent-inventory.mdx) and associate it with a project.
5. Review what it is configured with in [Agent Configuration](./agent-configuration.mdx) and what it sends in the [Audit Trail](./gateway-audit-trail.mdx).
6. Add controls with [Gateway Policies](./gateway-policies-and-plugins.mdx), [OPA/Rego Policies](./opa-rego-policies.mdx), and [Policy as Code](./policy-as-code.mdx).

## Related

- [Gateway Basics](./gateway-how-it-works.mdx) — one request followed from Viper through the gateway and back
- [Agent Onboarding](./agent-onboarding.mdx) — the steps from an empty machine to a signed, audited agent
- [Associated Agents](./associated-agents.mdx) — link an agent to a project so project policy applies
- [Plugin Configurations](./plugin-configurations.mdx) — the plugins in effect for a project and their activation history
- [Roles and Permissions](./roles-and-permissions.mdx) — who can register, associate, and revoke agents
- [LLM Gateway Overview](../llm-gateway/overview.mdx) — the gateway service, for operators and integrators