ovos-persona-server

A single HTTP server that exposes one or more OVOS Personas as eight concurrent API surfaces — so any LLM client (OpenAI SDK, LangChain, Ollama tools, Anthropic SDK, Google Gemini SDK, Cohere SDK, HuggingFace TGI client, AWS Bedrock client, or any A2A agent) can talk to your OVOS persona without changes.

Table of Contents

• What is a Persona?
• Installation
• Quick Start
• Serving Several Personas
• API Surfaces
• A2A Endpoint
• Persona Config Examples
• Streaming
• RAG: Files & Vector Stores
• Embeddings
• Authentication
• Troubleshooting

What is a Persona?

An OVOS Persona is a JSON file that chains together one or more handler plugins (also called solvers, the legacy key name). Handlers are tried in order until one returns an answer. You can mix LLMs, knowledge bases, and fallback bots in a single persona — no GPU required for non-LLM setups.

See the OVOS technical manual: Personas for the full persona file schema.

{
  "name": "OldSchoolBot",
  "handlers": [
    "ovos-solver-wikipedia-plugin",
    "ovos-solver-ddg-plugin",
    "ovos-solver-plugin-wolfram-alpha",
    "ovos-solver-wordnet-plugin",
    "ovos-solver-rivescript-plugin",
    "ovos-solver-failure-plugin"
  ],
  "ovos-solver-plugin-wolfram-alpha": { "appid": "YOUR_API_KEY" }
}

Find solver plugins at github.com/OpenVoiceOS.

Installation

# Base server (no A2A)
pip install ovos-persona-server

# With A2A server support
pip install 'ovos-persona-server[a2a]'

With uv (recommended in OVOS workspaces):

uv pip install 'ovos-persona-server[a2a]'

Quick Start

# Start serving a persona on port 8337
ovos-persona-server --persona /path/to/my-persona.json

# Also expose it as an A2A agent
ovos-persona-server \
  --persona /path/to/my-persona.json \
  --a2a-base-url http://localhost:8337/a2a

The server binds to 0.0.0.0:8337 by default. Visit http://localhost:8337/docs for the interactive API reference (Swagger UI).

Serving Several Personas

One process can host several personas. A persona's name is its model id: clients select one with the model field of whichever API they speak.

ovos-persona-server \
  --persona /path/to/assistant.json \
  --persona /path/to/rivescript-bot.json

# or load a whole directory
ovos-persona-server --personas-dir /etc/ovos/personas --default-persona assistant

GET /openai/v1/models (and GET /ollama/api/tags) list every loaded persona. An unknown model returns HTTP 404 listing the available names. With a single persona the model field stays advisory, so existing deployments are unaffected.

model selects a persona — a solver chain, a system prompt, and memory settings — not the LLM behind it. To change the LLM, edit the persona JSON.

Per-surface behaviour, the default-persona rule, and how state is kept separate: docs/multi-persona.md.

API Surfaces

Every API is served on a vendor-prefixed path so multiple clients can coexist without conflict.

Deprecated legacy paths

For backwards compatibility, /v1/... maps to /openai/v1/... and /api/... maps to /ollama/api/.... These paths send Deprecation and Link response headers and will be removed in a future major version. Migrate to the prefixed paths.

Quick test with curl

# OpenAI-compatible chat
curl -s http://localhost:8337/openai/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"","messages":[{"role":"user","content":"hello"}]}' \
  | python3 -m json.tool

# Ollama-compatible chat
curl -s http://localhost:8337/ollama/api/chat \
  -H 'Content-Type: application/json' \
  -d '{"model":"","messages":[{"role":"user","content":"hello"}]}'

A2A Endpoint

ovos-persona-server can expose your persona as a standard A2A agent server, enabling any A2A client to interact with it — including ovos-a2a-agent running on another OVOS instance.

Enable A2A

ovos-persona-server \
  --persona my-persona.json \
  --a2a-base-url http://myhost:8337/a2a

The --a2a-base-url flag:

• Activates the A2A endpoint at /a2a.
• Sets the url field in the Agent Card returned at GET /a2a/.well-known/agent.json.
• Must be the publicly reachable URL of the /a2a mount — this is what A2A clients use to discover the server.

Verify

# Fetch the Agent Card
curl http://localhost:8337/a2a/.well-known/agent.json | python3 -m json.tool

# Send a message
curl -X POST http://localhost:8337/a2a/ \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0",
    "id": "1",
    "method": "message/send",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"kind": "text", "text": "hello"}]
      }
    }
  }'

Connecting ovos-a2a-agent to this server

On another OVOS instance:

{
  "name": "remote-persona",
  "handlers": ["ovos-a2a-agent"],
  "ovos-a2a-agent": {
    "url": "http://myhost:8337/a2a"
  }
}

A2A streaming

The A2A endpoint supports message/stream. Persona sentence chunks are emitted as TaskArtifactUpdateEvent SSE events. Enable streaming on the client side (e.g. "streaming": true in ovos-a2a-agent config).

A2A without a2a-sdk

If a2a-sdk is not installed and --a2a-base-url is provided, the server starts normally and logs a warning. All other API surfaces continue to work.

Persona Config Examples

LLM persona (OpenAI-compatible backend)

{
  "name": "gpt-persona",
  "handlers": ["ovos-openai-plugin"],
  "ovos-openai-plugin": {
    "api_key": "sk-...",
    "model": "gpt-4o-mini"
  }
}

Knowledge-base + LLM fallback

{
  "name": "smart-assistant",
  "handlers": [
    "ovos-solver-wikipedia-plugin",
    "ovos-solver-ddg-plugin",
    "ovos-solver-wordnet-plugin",
    "ovos-openai-plugin",
    "ovos-solver-failure-plugin"
  ],
  "ovos-openai-plugin": {
    "api_key": "sk-...",
    "model": "gpt-4o-mini"
  }
}

Rivescript chatbot (no GPU, no API key)

{
  "name": "rivescript-bot",
  "handlers": [
    "ovos-solver-rivescript-plugin",
    "ovos-solver-failure-plugin"
  ]
}

Streaming

All seven non-A2A APIs support SSE streaming where the upstream spec defines it. Pass "stream": true (OpenAI / Cohere / TGI) or the equivalent for each API. See docs/streaming.md for per-API details.

OPM Tool Plugins — MCP and UTCP exposure

Installed ToolBox plugins (OPM entry-point group opm.agents.toolbox) are automatically surfaced over two protocols when the server starts.

Installing the MCP extra

pip install ovos-persona-server[mcp]

Without the [mcp] extra only the UTCP endpoints are active.

UTCP — Universal Tool Calling Protocol

Two endpoints are added at /tools:

Fetch the manual:

curl http://localhost:8337/tools/manual

Response shape:

{
  "utcp_version": "1.0",
  "tools": [
    {
      "name": "my_tool",
      "description": "Does something useful.",
      "tool_provider": {
        "type": "http",
        "method": "POST",
        "url": "http://localhost:8337/tools/my_tool",
        "content_type": "application/json"
      },
      "inputs": [
        {"name": "query", "type": "string", "required": true, "description": "Search query"}
      ],
      "output_schema": { ... }
    }
  ]
}

Invoke a tool:

curl -X POST http://localhost:8337/tools/my_tool \
     -H "Content-Type: application/json" \
     -d '{"query": "hello"}'

MCP — Model Context Protocol

When the [mcp] extra is installed, the server mounts an MCP SSE endpoint at /mcp. Each installed ToolBox tool is registered as an MCP tool with the name, description, and JSON Schema derived from its OPM definition.

Claude Desktop / MCP client config:

{
  "mcpServers": {
    "ovos-persona-tools": {
      "url": "http://localhost:8337/mcp/sse"
    }
  }
}

Standalone stdio MCP server (for clients that spawn a subprocess):

ovos-persona-tools-mcp

This runs the same tool set over the stdio MCP transport.

Writing a ToolBox plugin

Implement ToolBox from ovos_plugin_manager.templates.agent_tools and register it under the opm.agents.toolbox entry-point group:

# pyproject.toml
[project.entry-points."opm.agents.toolbox"]
my_toolbox = "my_package.toolbox:MyToolBox"

The server picks it up automatically on the next start.

Client side usage

The OpenAI and Ollama routers expose /embeddings endpoints. These require a solver plugin that implements get_embeddings(text). If no such solver is loaded the endpoint returns HTTP 501. See docs/embeddings.md.

Authentication

The server itself does not enforce authentication — deploy behind a reverse proxy (nginx, Caddy, Traefik) with TLS and auth if public exposure is required. For the A2A endpoint, A2A clients that require bearer tokens can be configured on the client side (api_key in ovos-a2a-agent config).

Troubleshooting

Failed to load persona (500 on startup) The persona JSON file was not found or is invalid. Check the --persona path and validate the JSON.

404 model_not_found on chat requests Several personas are loaded and the model you sent is not one of them. GET /openai/v1/models lists the valid names. See docs/multi-persona.md.

duplicate persona name on startup Two loaded persona files declare the same name. Names are the model ids and must be unique.

All requests return 500 Persona chat failed The underlying solver chain failed. Check solver plugin installation and their individual configs (API keys, model paths, etc.).

A2A endpoint not available after starting with --a2a-base-url a2a-sdk is not installed. Install it:

uv pip install 'ovos-persona-server[a2a]'

Then restart the server.

Embeddings return 501 No solver with get_embeddings() is loaded and no embeddings plugin could be loaded. Configure TEXT_EMBEDDINGS_PLUGIN or add an embeddings solver to the persona's solvers list.

Legacy /v1/ paths return responses with Deprecation header This is expected. Migrate to /openai/v1/ paths. See docs/deprecation.md.

RAG: Files & Vector Stores

ovos-persona-server exposes an OpenAI-compatible Retrieval-Augmented Generation surface: upload documents, embed and index them, search by similarity, and feed the results into any chat endpoint. Files, embedding, and the vector DB are all backed by swappable OVOS plugins. Full reference: docs/rag.md; runnable scripts in examples/.

Drop-in OpenAI replacement: any third-party app built on OpenAI's Files / Vector Stores / Embeddings endpoints can point at this server by changing only its base_url — a self-hosted, private, zero-cost RAG backend with no code changes.

uv pip install 'ovos-persona-server[rag]' ovos-gguf-plugin ovos-chromadb-embeddings-plugin

TEXT_EMBEDDINGS_PLUGIN=ovos-gguf-embeddings-plugin EMBEDDINGS_MODEL=all-MiniLM-L6-v2 \
EMBEDDINGS_DB_PLUGIN=ovos-chromadb-embeddings-plugin \
ovos-persona-server --persona examples/persona_rag.json

Drive it with the official openai SDK — files hit /openai/v1/files, the store is a collection in the configured vector DB:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8337/openai/v1", api_key="unused")

f = client.files.create(file=("cats.txt", b"cats are fluffy animals that sit on mats."),
                        purpose="assistants")
store = client.vector_stores.create(name="kb")
client.vector_stores.files.create(vector_store_id=store.id, file_id=f.id)

hits = client.vector_stores.search(vector_store_id=store.id, query="fluffy animal", max_num_results=3)
print([(h.file_id, h.score) for h in hits.data])

For conversational use, the companion ovos-openai-plugin ships PersonaServerRAGMemory — a persona memory plugin that searches a vector store and injects the retrieved context, composing with any chat backend. See docs/rag.md and examples/rag_memory_plugin.py.

Backend vs hosted agent. By default the chat endpoints are a stateless backend (CHAT_MEMORY=off): the client owns conversation state and drives the Files / Vector-Stores endpoints itself — the correct behaviour for a drop-in OpenAI replacement or any multi-user deployment. Set CHAT_MEMORY=transparent to run a single-user hosted agent where the server folds the persona's memory_module (history + RAG) into every turn and persists it per session. See docs/rag.md.

Embeddings

A single, swappable embeddings service backs every vendor surface — OpenAI (/openai/v1/embeddings), Ollama (/ollama/api/embed and /embeddings), Cohere (/cohere/v1/embed), Gemini (:embedContent / :batchEmbedContents), HuggingFace TGI (/tgi/embed), AWS Bedrock (Titan/Cohere embed models via /invoke), and the vector-store search path all delegate to the same backend. (Anthropic has no first-party embeddings API, so it has no embed endpoint.) This mirrors how inference is backed by one shared persona: swap the embeddings provider once and it changes everywhere. Per-surface request/response shapes: docs/embeddings.md.

The backend is any OVOS text-embeddings plugin, configured through the environment:

Point TEXT_EMBEDDINGS_PLUGIN at a local model (the default gguf plugin) or at any remote embeddings API via an OpenAI-compatible plugin and the matching EMBEDDINGS_URL / EMBEDDINGS_MODEL. When no embeddings plugin is available the server falls back to a persona solver exposing get_embeddings.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8337/openai/v1", api_key="")
resp = client.embeddings.create(model="", input=["hello", "world"])
print(len(resp.data), "vectors")

from ollama import Client

client = Client(host="http://localhost:8337/ollama")
print(client.embed(model="", input=["hello", "world"]).embeddings)

Credits

Developed by TigreGótico for OpenVoiceOS.

This project was funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429.