Skip to content
Download MarkdownBack to AethisJoin the waitlist

What is Aethis?

Aethis is a governance and control layer for AI agents. It sits between an agent and the systems it can affect, evaluates each action against policy, and records the decision and outcome together.

Your agent keeps running in your own application. Aethis governs the calls you route through it; it does not host your agent or govern calls that bypass it.

The governance loop

Section titled “The governance loop”
  1. Your agentMakes a tool call
  2. Aethis endpointIdentifies the Agent
  3. Policy checkEvaluates the exact action
  4. Your MCP serverReceives permitted calls in Enforce mode
  1. Identify the caller. Your MCP client calls the Agent’s Aethis endpoint and sends its access key. The key identifies the Agent; it is not a credential for the upstream server.
  2. Resolve the destination. Aethis checks the Agent’s assigned Connections and the discovered tool. Attaching a Connection establishes a path; policy determines what may happen on it.
  3. Evaluate the action. The applicable Policies evaluate the tool, target, and arguments. The policy decision is Allow, Require Approval, or Block.
  4. Apply the mode. In Enforce, permitted calls reach the upstream server, blocked calls stop, and calls requiring approval are held. In Observe, Aethis records the decision while execution continues unchanged.
  5. Keep the evidence. The Action Record connects the request, policy version, any review, and execution outcome.

A policy decision and an execution outcome answer different questions. Allow says the action was permitted. Execution succeeded says it ran successfully. Read both before concluding what happened.

The objects you work with

Section titled “The objects you work with”
Object What it answers Where to start
Agent Who is making the call, and who owns it? Register an Agent
Connection Which upstream system may this Agent reach? Connect an MCP server
Policy Which actions may run, need review, or must stop? Create a Policy
Approval request May this exact action proceed once? Review an approval
Action Record What was requested, decided, approved, and executed? Investigate in Activity

Agents and Connections belong to a workspace. Use separate workspaces when environments need separate credentials, access, or ownership. Organization settings govern the wider team. See Organizations and workspaces.

Follow a refund example

Section titled “Follow a refund example”

Suppose a Billing Agent uses a published refund policy with these boundaries:

  • Up to SGD 500: Allow.
  • Above SGD 500 and up to SGD 1,000: Require Approval.
  • Above SGD 1,000: Block.

For a SGD 750 refund, the policy returns Require Approval. In Enforce mode, the call is held for review. A reviewer checks the account, amount, target, and arguments, then chooses Approve Once or Deny, with a reason.

Approval does not itself execute the refund. It authorizes one exact action. Changing the amount or target requires another evaluation. Follow the request into its Action Record to see whether execution actually happened.

In Observe mode, the same rule can return Require Approval without holding execution. An observed decision is evidence of what the policy said, not proof that the action was stopped.

Use Verify a governance boundary to check a policy with safe example inputs before relying on it.

Integration paths

Section titled “Integration paths”

MCP tool calls. Your MCP-compatible client uses one Aethis endpoint per Agent, regardless of how many Connections are attached:

URL:     https://mcp.aethis.io/v1/agents/{agent_id}/mcp
Header:  Authorization: Bearer <aethis_agent_token>

Copy the values from the dashboard and keep the access key in a secret manager. No SDK is required. Your remote MCP servers stay where they are.

Model calls. Direct model routing supports OpenAI chat completions through a Model Route. A Model Route and an MCP Connection are different destinations; choose the path that matches your client’s request. See Connect an MCP server and Model Routes.

Who is Aethis for?

Section titled “Who is Aethis for?”

Teams that give agents access to consequential tools: payments, customer records, deployments, or internal systems. Builders configure the boundary; reviewers decide exceptional actions; operators inspect evidence and resolve problems.

How is Aethis different from an AI gateway or observability product?

Section titled “How is Aethis different from an AI gateway or observability product?”

A gateway routes requests. Observability helps you inspect what happened. Aethis adds an action-level policy decision, an approval when required, and evidence connecting the decision to the outcome. These categories can overlap; Aethis’s focus is governing the action in the request path.

Next steps

Section titled “Next steps”