Skip to main content

How agents are exposed

An agent you build in Orchestrate can be reached in several ways. Each one runs the agent's live version, goes through the same approval gate, policy checks and guardrails, and shows up in Runs & traces. They differ in who is calling and how they prove who they are.

Who callsHowCredentialRead
Your own applicationThe Agent API: OpenAI-compatible chat completions, async runs with webhooks, server-held threadsA project API key (vcpj_)Agent API
People in a browserA hosted chat page, or a widget embedded in your own siteTheir Orchestrate sign-in, or nothing when anonymous chat is onHosted chat and widget
Another agentA2A (the Agent2Agent protocol), or an agent given another agent as a toolAn agent key (vcag_), or the calling agent's own runAgents calling agents
People in Microsoft Teams, Slack or WhatsAppA channel connected to the agentThe vendor's signed webhookChannels
People on the phoneA phone number on a SIP trunkThe carrier's trunkPhone calls (SIP)
People in the consoleThe Playground, voice includedTheir Orchestrate sign-inPlayground

Code agents and console-made agents are reached the same way. Where a surface behaves differently for a code agent, its page says so.

The Connect section​

Every agent's page has a Connect section. On a code agent it is the Connect tab. It shows, for that agent:

  • the Agent API address, https://<your API host>/v1/agents/<agent name>, and a button to create a project API key restricted to this agent
  • copyable snippets for chat completions, threads and async runs, in cURL, Python and TypeScript, and for calling the agent from LangGraph, the OpenAI Agents SDK, LlamaIndex, CrewAI and ADK
  • End-user chat: switch on the hosted page and widget, and copy the page link and the embed snippet
  • Channels: connect Teams, Slack or WhatsApp
  • A2A: the agent's card address, who may call it, client snippets, the agent's identity and sponsor, and the MCP server URL for outside MCP clients. See Agents calling agents.

To start an agent on a schedule or from a signed webhook, see Schedules and webhooks. How the agent runs on a voice or phone call is set in the agent's Voice section. See Voice agents.

If the agent has no live version, the section says so. Nothing can call an agent until a version is live. See Live versions and the release gate.

Who can switch things on​

Each surface puts the agent in front of people who are not members of its project, so turning one on needs the admin role on the project:

  • creating a project API key or an agent key
  • switching on the hosted chat, and above all anonymous chat
  • connecting a channel
  • choosing which agents may call an agent over A2A

Developers can read the settings and use the snippets.

What every surface shares​

  • The live version answers. Making another version live changes the answer on every surface at once.
  • Approvals wait. A tool call that needs a person waits on every surface. The caller is told it is waiting, and the answer arrives when someone decides. See Human approval.
  • Usage controls apply. Budgets, rate limits and key scopes are checked before a run starts. See Usage controls.
  • Stored content is redacted. What a caller gets back in flight is the model's answer. What Orchestrate keeps (run records, thread history, traces) goes through the PII redactor first. See PII guardrails.
  • Feedback. People can rate answers with a thumb up or down on every surface. See Online evals and feedback.
Was this page helpful?
On this page
Talk to our teamQuestions about licensing or sizing VoisX Orchestrate for your site? We can help.Talk to sales →