# Thoth overview

> Decide what your AI agent may do before it acts, then inspect the decision.

Source: https://docs.atensecurity.com/docs

Thoth, from Aten Security, governs AI agent actions at the moment they execute. The SDK and proxy put that check between an agent and its tools.

## The action loop

Flow: Agent proposes a tool call (Who is acting, on what, and for whom?) → Thoth evaluates policy (ALLOW · STEP_UP · BLOCK) → Execute only when permitted (Inspect the decision and available evidence)

`ALLOW` permits execution. `STEP_UP` waits for an authorized approval.
`BLOCK` refuses the action. Keep the tool behind the governed wrapper so a
model cannot bypass that decision by calling the original function.

## Start with one agent

- [Your first governed agent](https://docs.atensecurity.com/docs/getting-started/quickstart): Install a tested SDK, run a harmless read, and verify a refusal. No model API key needed.
- [Instrument your own agent](https://docs.atensecurity.com/docs/getting-started/your-agent): Wrap real callbacks, attribute actions, and handle refusals safely.

Need a tenant first? See [Trial access and setup](https://docs.atensecurity.com/docs/getting-started/trial-quickstart).
Already integrating? Use the [SDK references](https://docs.atensecurity.com/docs/sdks) or [REST API reference](https://docs.atensecurity.com/docs/api-reference).

## Choose another integration

| Your task | Start here |
| --- | --- |
| Govern an agent written in Python, TypeScript, or Go | [Install an SDK](https://docs.atensecurity.com/docs/getting-started) |
| Govern external MCP tools used by Claude Desktop | [MCP proxy](https://docs.atensecurity.com/docs/mcp/proxy) |
| Configure existing governance settings as code | [Terraform](https://docs.atensecurity.com/docs/terraform-provider) or [Pulumi](https://docs.atensecurity.com/docs/pulumi-provider) |
| Manage an enterprise endpoint fleet | [Deployment guides](https://docs.atensecurity.com/docs/deployments) |
| Reconcile tenant settings in Kubernetes | [Kubernetes operator](https://docs.atensecurity.com/docs/kubernetes-operator) |
| Investigate or recover a deployment | [Operations](https://docs.atensecurity.com/docs/operations) |

The MCP proxy checks other servers' tool calls; it does not expose Thoth
configuration as MCP tools. Use the configuration providers or
[control-plane CLI/API](https://docs.atensecurity.com/docs/headless-gitops) for supported administrative tasks.

## Understand what you verified

A policy decision, a completed tool call, a behavioral event, and a signed
receipt are different records. Inspect the reason, actor context, and
correlation identifiers in your environment. Verify receipt signatures and
available exports before treating them as evidence for an external review.
A `user_id` supplied by your application is attribution context; it does not
by itself prove a human authorized an action.

Read [Core concepts](https://docs.atensecurity.com/docs/getting-started/concepts) for scope, sessions,
approvals, and failure behavior. SDKs, proxies, and infrastructure providers
have separate release versions; pin and validate the releases you use.