Skip to content
G

Guardyx AI

Get started

Autonomous AI agents under your control

Guardyx AI is the Runtime Control Layer for Autonomous AI, for developers and anyone whose AI agents act on real accounts.

Your agents send email, push code, and spend money with your credentials. Guardyx checks each action against your rules first, holds the risky ones for you, and records every decision.

Every AI action passes through Guardyx.

Get startedRead the docs

Approval requested

Held for a human before it runs.

High riskPending
Agent
Claude Desktop
Tool
GMAIL_SEND_EMAIL

Payload

{
  "to": "ALL-CUSTOMERS@list.acme.com",
  "body": "Effective immediately, new Q3 pricing…"
}

Policy

Bulk external email to a distribution list — approval required.

Your AI agents aren't just chatting anymore.

Claude, Cursor, and the agents you build now act with your credentials — faster than you can read what they're about to do.

If your agents only draft text and never call a real tool, you probably don't need this yet.

It only takes one.

  • Force-push over your main branch.
  • Email everyone in your contacts.
  • Delete files or a table you still needed.
  • Paste a secret key somewhere public.
  • Spend money with your card or API credits.

Each happens in seconds, under your name, and you only find out afterward.

All your MCPs in one place.

Point every agent and MCP client at Guardyx AI. One control point, one record.

Your agents

  • CrewAI
  • Microsoft AutoGen
  • Google ADK
  • OpenAI Agents SDK
  • Anthropic Claude Agent SDK
  • LangGraph & LangChain
  • OpenAI tool calling
  • ChatGPT & Claude desktop, over MCP

Guardyx AI

Policy · Approvals · Audit

hosted, or in your VPC

The systems they act on

  • Sales & email tools
  • Payments & ledgers
  • Production infrastructure
  • Internal data & records

Connect your MCP clients.

Claude Desktop, Cursor, and the Claude.ai connector sign in over OAuth 2.1. Connect directly to ChatGPT or Claude desktop applications — no adapter needed.

Bring the agents you already run.

Native adapters for CrewAI, AutoGen, Google ADK, the OpenAI and Anthropic agent SDKs, and LangGraph, with LangChain tools governed through that same LangGraph adapter.

Connectors included.

One bill for the tools your agents call and the control over them.

See how agents connect

Control at the moment of action.

Not a log you read afterward. The decision happens before the action runs.

Your agent proposes

GMAIL_SEND_EMAIL

to every customer on a list

Guardyx AI

Checks it against your policy

before anything runs. If the check fails, nothing runs.

  • Allowed

    In policy, so it runs now.

  • Held

    Waits for a human to approve.

  • Denied

    Never reaches the system.

Audit trail

Every decision recorded

tamper-evident, including the ones that never ran.

Fails closed.

Guardyx AI fails closed by default: if the policy check cannot complete, the action does not run.

Keeps the receipts.

A tamper-evident record of what was allowed, held, or blocked.

Runs where you need it.

Hosted, or self-hosted in your own VPC or on-prem.

This is the screen you'll use.

The real approvals queue: risky tool calls held before they run.

The Guardyx AI approvals queue: pending agent tool calls such as GMAIL_SEND_EMAIL and NOTION_CREATE_PAGE, each tagged high or critical risk and held for a human to review before they run.
An approval's detail panel in Guardyx AI: the agent, the tool it called, the request payload, and the policy that required review, with approve and deny actions and the recorded outcome.

Open a held action, see why it stopped, approve or deny.

Batch review in Guardyx AI: several low-risk email actions selected together, with controls to approve or deny all of them at once.

Clear routine actions in one batch.

Prevent. Review. Prove.

Nobody wants to be the person explaining why an agent did something nobody approved. Guardyx AI does all three jobs, on every action.

  • Prevent

    No risky action runs without a policy decision.

    Intercept the action and check it against your policy.

  • Review

    You see only what matters.

    Hold the exceptions for your sign-off; let routine work through.

  • Prove

    Export the evidence when you need it.

    Run it, and log the decision as tamper-evident evidence.

The Guardyx AI Runs Explorer: every governed run with who started it, the tools it called, the policy decision, and the outcome — the record you review and export as evidence.

Every run and its decision, exportable as a tamper-evident evidence pack.

See the full platform

Bringing it to your company?

The same control layer runs company-wide: self-hosted in your VPC, with SSO, SAML and SCIM, and evidence your auditor can read.

Request early access to Enterprise

Start free in your browser.

Every AI action passes through Guardyx.

5,000 governed calls a month, connectors included, for up to 3 teammates. Every governance feature. No card, no expiry.

Start freeSee pricingRead the docs