Agents and Connections

Two objects describe what an AI agent can reach: the Agent, which says who is calling, and the Connection, which says where the call may go.

Agents

An Agent is the identity Aethis associates with a governed call. Each Agent belongs to one workspace and has:

• a name and an Owner (the team accountable for it)
• an Agent identity such as agent_billing_prod
• one Aethis MCP server URL, https://mcp.aethis.io/v1/agents/{agent_id}/mcp
• an Aethis access key, the Agent’s scoped credential, which your client sends as Authorization: Bearer <aethis_agent_token>; the dashboard shows only its Agent key fingerprint after creation
• a setup type: MCP-only, Model-only or Combined
• assigned Connections, Policies and basic controls

A connection or a key does not grant permission on its own. Having a Connection makes a tool reachable; the Agent’s Policies decide whether a particular call may run.

Agent sections

Every Agent has four sections:

Section	Shows
Overview	Usage & limits (including the spend budget), Agent details, the access key fingerprint, and Basic controls.
Connections	The Connections assigned to this Agent and their status.
Policies	The Policies that apply to this Agent, whether assigned directly or required by the organization.
Tools	Every tool the Agent can reach, the policies that apply to it, and its current access.

Basic controls

Basic controls on the Agent’s Overview apply to every call the Agent makes:

• Data protection: PII redaction, Payment data redaction and Secrets redaction mask sensitive values before they leave for a provider or tool. Redaction is an Agent control, not a policy decision.
• Approvals & evidence: Unknown tool approval holds untrusted tools for review, and Tamper-evident evidence records a hash-chained decision and policy snapshot.

Spend budgets

Spend is visible at both Agent and workspace level:

• Agent (Usage & limits on the Agent’s Overview): Monthly spend cap, Spend this month and an Alert at percentage. The Agent figures track the workload against its budget; Alert at is a warning threshold and does not stop execution.
• Workspace (Workspace spend controls in workspace settings): an aggregate Monthly spend cap and Current spend. This is the enforced monetary ceiling for supported cost-bearing execution in the workspace.

Do not assume the Agent’s Monthly spend cap independently stops a call. Use it to track and alert on the Agent’s contribution, and use the workspace cap as the documented execution boundary.

A spend budget is separate from your plan’s allowance of governed actions. See Manage spend and usage or Set a spend budget.

Agent status

An Agent is Active, Degraded, Paused, Suspended or Disabled. Paused is a temporary stop you start with Pause Route: governed routes stop, the configuration remains, and Resume Route restores it. Suspended is a protective stop enforced by Aethis, which lasts until the triggering condition is resolved. Disabled means turned off administratively until someone re-enables it. See Statuses.

Connections

A Connection is an upstream destination an Agent may use through Aethis.

• MCP Connections point at a remote MCP server by its HTTPS URL, with Bearer, OAuth, mTLS or None authentication. This is the main path in the beta.
• Model Routes send model requests to a model provider through the Aethis Model Gateway. Direct model routing supports OpenAI (chat completions) only.

A Connection belongs to one workspace, and you can assign it to any Agent in that workspace. Everything Aethis knows about the server’s tools comes from tool discovery.

The execution boundary

Each MCP Connection has an execution boundary that all assigned Agents share:

• Governed: calls can be evaluated and forwarded.
• Frozen: execution is paused, usually because the server’s tools changed and the change hasn’t been reviewed. Calls through the Connection don’t run until someone accepts or rejects the change.

One endpoint per Agent

Your client configures one Aethis MCP endpoint and one credential per Agent, however many Connections the Agent has:

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

The URL isn’t a secret; the access key is. Behind that one endpoint, Aethis resolves each call in order:

Agent  ──►  assigned Connections  ──►  discovered capabilities  ──►  effective policy  ──►  upstream MCP server

Attaching or detaching a Connection changes what the Agent can reach through the same endpoint. You don’t reconfigure the client.

How the pieces fit

Workspace
 └─ Agent  (identity + one MCP endpoint + access key + basic controls)
     ├─ Connections   (which MCP servers or OpenAI routes it can reach)
     ├─ Policies  (what it may do there: Allow, Require Approval, Block)
     └─ Tools         (the discovered tools, with the access that results)

The Control Map draws this picture for any Agent.