AI Agents: MCP, UTCP and A2A¶

In a nutshell

OpenVoiceOS (OVOS) can give its speech, translation and reasoning abilities to outside AI agents, and it can use tools and agents that live outside OVOS. Three open protocols make this work: MCP, UTCP and A2A. Each page in this section is one task that you can do, with the exact steps and the real output.

What you can do¶

What you can do	Task	Component	Component docs
Transcribe speech from Claude Code, Claude Desktop, opencode, Mistral Vibe or any MCP or UTCP agent	Transcribe audio from an AI agent	`ovos-stt-server`	AI agent integration
Make speech from an AI agent	Let an AI agent speak	`ovos-tts-server`	Agent integration
Translate text and detect its language from an AI agent	Translate text from an AI agent	`ovos-translate-server`	AI agent integration
Let other agents talk to an OVOS persona	Talk to an OVOS persona over A2A	`ovos-persona-server`	A2A endpoint
Write a tool that every OVOS agent component can use	Write a tool plugin	`ovos-plugin-manager`	Agents
Call the tools of a persona from an AI agent	Use persona tools over MCP and UTCP	`ovos-persona-server`	API surfaces
Share tools with every program on an OVOS device	Share tools on the messagebus	`ovos-PHAL-plugin-tools`	Bus API reference
Give a persona tools from an MCP or UTCP server	Give a persona MCP and UTCP tools	`ovos-agentic-loop`, `ovos-tool-adapters`	Toolboxes, MCP, UTCP
Let a persona answer through an agent on another server	Answer through a remote A2A agent	`ovos-a2a-agent-plugin`	Configuration
Connect HiveMind voice satellites to an A2A agent	Connect a HiveMind satellite to an A2A agent	`hivemind-a2a-agent-plugin`	Configuration

The three protocols¶

MCP (Model Context Protocol) lets an AI assistant, such as Claude Code, Claude Desktop, opencode or Mistral Vibe, find and call the tools of a server.
UTCP (Universal Tool Calling Protocol) gives an agent a JSON manual at GET /utcp. The manual tells the agent how to call the server's own HTTP endpoints, so no wrapper process is necessary.
A2A (Agent-to-Agent) lets one agent give a full question to another agent and get the answer back.

Protocol	How an agent finds the tools	How an agent calls a tool	What the server needs
UTCP	`GET /utcp` returns a UTCP 1.0 manual	The agent calls the HTTP endpoints that the manual names	Nothing. The endpoint is always on.
MCP	MCP `initialize` and `tools/list` at `/mcp`	MCP `tools/call` over Streamable HTTP	The `mcp` extra and the `--mcp` flag
A2A	`GET /a2a/.well-known/agent-card.json`	JSON-RPC 2.0 `SendMessage`	The `a2a` extra and the `--a2a-base-url` flag

flowchart LR
    subgraph Outside OVOS
        Agent["AI agent<br/>Claude Code, Claude Desktop,<br/>opencode, Mistral Vibe,<br/>UTCP or A2A client"]
        ExtTools["MCP or UTCP server"]
        ExtA2A["A2A agent"]
    end
    subgraph OVOS
        STT[STT server]
        TTS[TTS server]
        TX[Translate server]
        PS[Persona server]
        Loop["Persona with an<br/>agentic loop"]
        Solver["Persona with<br/>ovos-a2a-solver"]
        HM["HiveMind master with<br/>hivemind-a2a-agent-plugin"]
    end
    Agent -->|"MCP, UTCP"| STT
    Agent -->|"MCP, UTCP"| TTS
    Agent -->|"MCP, UTCP"| TX
    Agent -->|"MCP, UTCP, A2A"| PS
    Loop -->|"MCP, UTCP"| ExtTools
    Solver -->|A2A| ExtA2A
    HM -->|A2A| ExtA2A

Diagram: On the left, outside agents call the four OVOS servers. On the right, three OVOS components call tools and agents outside OVOS.

Default ports¶

Server	Command	Default port
STT	`ovos-stt-server`	`8080`
TTS	`ovos-tts-server`	`9666`
Translate	`ovos-translate-server`	`9686`
Persona	`ovos-persona-server`	`8337`

Security¶

The agent endpoints have no authentication

/mcp, /utcp, /tools/* and /a2a accept every request. Any client that can reach the port can call every tool. Bind the servers to 127.0.0.1 or to a trusted network. To serve other hosts, put a reverse proxy with authentication in front of the server.

Read next: Transcribe audio from an AI agent Related: Agent Tool Plugins · Agentic Loops · Persona Server