Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.
Version: 11.3.0

Anthropic Cloud Agents

Overview​

An Anthropic cloud agent is an agent HYPR AgentPass™ creates and hosts for you on Anthropic's Managed Agents platform. You choose its name, model, system prompt and tools in the AgentPass console. AgentPass provisions the agent on Anthropic and keeps the two in step from then on.

The agent runs at Anthropic, but its tools live at your gateway. Every MCP tool you give it is a gateway connector. Each tool call therefore comes back to AgentPass, where it is policy-checked, approved where required and logged before anything happens.

Before You Begin​

You need two things before you can create an agent:

  • An Anthropic Managed Agents connector. It holds the account and API token that AgentPass authors and runs agents on. Setup follows.
  • A gateway address Anthropic can reach. The agent's tools are URLs on your gateway, called from Anthropic's infrastructure, so your AgentPass deployment must be reachable at its public URL. If you are unsure, confirm with your HYPR representative.

Give the agent at least one MCP tool connector: an MCP server or a REST API connector. An agent with no tools can converse but cannot act. See Configuring Connectors.

Set Up the Anthropic Account and API Token​

The connector holds the account credentials. Every agent is pinned to the connector it was created on.

  1. In the console, open Connectors.
  2. Click Add Connector.
  3. Under Cloud Agents, choose Anthropic Managed Agents.
  4. Complete the fields described in the following table.
  5. Click Create Connector.
The Anthropic Managed Agents connector form with blank API-key, owner-email and webhook-signing-key fields

The connector form has these fields.

FieldWhat to enter
Anthropic Managed Agents API keyThe account's API key (sk-ant-…). Stored encrypted, and never shown again after saving
Owner emailThe supervisor who owns agents imported from this account
Webhook signing keyThe whsec_… secret from the Anthropic Console's Webhooks page. It lets AgentPass verify the session updates Anthropic sends back. Set it up unless you have a reason not to
API base URLLeave the default, unless you route through a proxy
Environment IDLeave blank. AgentPass resolves the execution environment automatically on first use
Treat the API key like the master key it is

This one token authors, runs and archives every agent on the account. Create it on a dedicated Anthropic account for AgentPass. To rotate it, save a new value on the connector. The field is write-only, so the new value replaces the old one.

Create an Agent​

  1. In Oversight → Cloud Agents, open the New agent menu.

  2. Choose New cloud agent.

  3. Choose a starter template. Templates prefill the prompt and a tool list for common jobs: Chief of Staff, Researcher, Recruiter, Sales Ops and Support Triage. Choose Custom to start blank. A template suggests only tools that exist as connectors on your deployment.

    The starter template picker for a new Cloud Agent
  4. Configure the agent. The fields are described in the following table.

    The Cloud Agent configuration form with its account, name, description, connector tools, roles, model and system prompt
  5. Save. AgentPass creates the agent on Anthropic and lists it in Cloud Agents in one step. If the Anthropic call fails, nothing is created on either side.

The configuration form has these fields.

FieldWhat it does
Anthropic Managed Agents accountThe connector the agent is authored on. Fixed once created
NameThe agent's display name
What this Cloud Agent doesA description shown wherever the agent is listed
MCP tools available to this Cloud AgentThe connectors the agent may call. Each row shows whether you have connected that tool yet
Session evaluatorTags the agent for Session Watch duty. This removes the provider sandbox tools
Policy authorTags the agent to draft policies on an administrator's behalf. The form warns This removes every tool from this agent: a policy author gets no sandbox, connector or gateway tools
Model (under Advanced)The Claude model the agent runs on
System prompt (under Advanced)The agent's persona and operating rules
Connect your tools once, and the agent reuses the connection

Tools that authenticate per person, such as OAuth-managed Slack or Google Workspace connectors, show a Connect button until you link your account. You connect once, and the agent reuses your saved connection at runtime. Each person the agent is assigned to connects their own accounts the same way, so the agent acts with that person's access, not yours.

How Tools Are Wired​

The tool list is part of the agent's definition. When you save, AgentPass pushes the agent's configuration to Anthropic. It includes one MCP server entry per selected connector, each a URL on your gateway. It also includes HYPR's own built-in tool server, which is how the agent requests approval when a policy asks for it.

Every one of those tool calls lands back at the gateway, so the gateway is the single place approval decisions are made. There is nothing to configure on the Anthropic side, and no second permission system to keep in step. Your policies decide each call, and the Request Log records it.

To change an agent's tools later, click Edit configuration on the agent's detail page. Saving updates the live Anthropic agent immediately.

The Provider Sandbox​

Besides your connectors, an Anthropic cloud agent normally carries Anthropic's built-in sandbox tools: shell, file operations and web fetch. They run inside Anthropic's environment.

The sandbox does not pass through the gateway

Sandbox tool use runs at the provider and never reaches AgentPass. No policy evaluates it, no approval gates it, and no audit row records it. Keep that in mind before you give an agent sensitive material. For this reason, agents tagged as session evaluators do not get the sandbox.

Talking to an Agent​

Open the agent, and click the Chat tab. A conversation with an Anthropic agent is a session. The agent works asynchronously and streams its replies, tool calls and status back as events. You can send follow-up messages, and interrupt a session that has gone off track.

When a tool call is held for approval, the agent is told and waits. It picks the task back up once a supervisor approves. The approval itself follows the standard AgentPass flow on the supervisor's phone.

The Chat tab of an Anthropic cloud agent: a request for a market briefing, the agent's call to a Jira tool through AgentPass, and its reply

Keeping AgentPass and Anthropic in Step​

Three actions keep the two sides consistent.

ActionWhat it does
SyncReconciles the whole Anthropic workspace into AgentPass. It imports agents created directly in the Anthropic console, and picks up remote edits. Administrator only
Re-sync (per agent)Pushes the agent's stored configuration back to Anthropic unchanged. Use it when the agent's tool URLs have gone stale, for example after your gateway's public address changes
Remote config (per agent)Shows the agent's record as Anthropic holds it, so you can compare it with the AgentPass side

Disabling an Agent​

To stop an agent, turn off the Enabled switch in the agent's header. Turning it off is reversible and immediate, and the gateway refuses the agent's tool calls until you turn it back on. See Enabling and Disabling.

Turning an Anthropic agent off also ends its sessions at Anthropic. Each running session is interrupted, and then archived, which makes it read-only. An archived session cannot be continued, even after you turn the agent back on, so people start a new chat instead. The agent's definition at Anthropic is untouched.

Example: A Chief of Staff Agent​

This example builds an agent from the Chief of Staff template.

  1. Create the agent. In Oversight → Cloud Agents → New agent → New cloud agent, choose Chief of Staff. Its suggested tools are your mail, calendar and Slack connectors. Select the ones you have, and click Connect on any that ask for your account.
  2. Test it. On the agent's Chat tab, ask "What's on my calendar today?" When the reply arrives, check the agent's Activity tab. Each call is attributed and policy-checked.
  3. Govern it. To keep a human in the loop for outbound actions, add a policy that requires approval for the mail connector's send tool. The agent requests approval and continues once it is granted. See Policies.
  4. Assign it. On the Assignments tab, assign the colleagues who should use it. Each person connects their own tool accounts on first use.
  5. Schedule it. Add a weekday-morning automation, such as "Summarize my unread mail and today's meetings," so the briefing is waiting before anyone asks. See Automating Cloud Agents.