mtpx 0.1.20

MTPX: Model Tool Protocol and agent runtime for Python

MTPX (Model Tool Protocol Extended)

MTPX is a protocol-first Python library for agent tool orchestration, built to support:

• Lazy tool loading by toolkit/category.
• Dependency-aware batch tool execution.
• Policy-aware execution based on tool risk.
• Multi-round model-tool-model loops.
• Provider adapters (now including Groq, Gemini, OpenAi, Anthropic, Openrouter, etc.).
• Transport primitives (stdio + HTTP + optional WebSocket envelope transport).
• Experimental MCP JSON-RPC adapter over the same runtime core.

Direction

This project has two explicit layers:

• MTP protocol: protocol entities and execution semantics.
• MTP Agent SDK: framework/runtime/providers/toolkits/transports built on top of MTP.

MCP support is an interoperability capability, not the product identity.

Canonical direction document:

• Project Direction

Quickstart

Install

From PyPI (recommended)

pip install mtpx

Common optional installs:

# Groq + dotenv helper
pip install "mtpx[groq,dotenv]"

# LM Studio local inference
pip install "mtpx[lmstudio]"

# Ollama local inference
pip install "mtpx[ollama]"

# OpenAI + Anthropic providers
pip install "mtpx[openai,anthropic,dotenv]"

# Web toolkits
pip install "mtpx[toolkits-web]"

# Database session stores
pip install "mtpx[stores-db]"

# Everything optional
pip install "mtpx[all]"

Verify installation:

python -c "import mtp; print(f'MTPX version {mtp.__version__} installed successfully!')"

From source (for development)

git clone https://github.com/yourusername/MTP.git  # Replace with your actual repo URL
cd MTP
python -m venv .venv
.venv\Scripts\activate  # On Windows
# source .venv/bin/activate  # On Linux/Mac
pip install -e .

Provider SDKs and dotenv (explicit alternative)

pip install "mtpx[groq,dotenv]"

Copy .env.example to .env and set your key:

GROQ_API_KEY=your_groq_api_key_here

Create an agent (local toolkits + Groq)

from mtp import Agent
from mtp.providers import Groq
from mtp.toolkits import CalculatorToolkit, FileToolkit, PythonToolkit, ShellToolkit

Agent.load_dotenv_if_available()

tools = Agent.ToolRegistry()
tools.register_toolkit_loader("calculator", CalculatorToolkit())
tools.register_toolkit_loader("file", FileToolkit(base_dir="."))
tools.register_toolkit_loader("python", PythonToolkit(base_dir="."))
tools.register_toolkit_loader("shell", ShellToolkit(base_dir="."))

provider = Groq(model="llama-3.3-70b-versatile")

agent = Agent.MTPAgent(
    provider=provider,
    tools=tools,
    instructions="Use tools when needed and return concise answers.",
    debug_mode=True,
    strict_dependency_mode=True,
)
response = agent.run("Calculate 25*4+10 and list files in current directory.", max_rounds=4)
print(response)

# Stream final response tokens:
agent.print_response("Give me a short summary.", max_rounds=4, stream=True)

# Stream structured runtime events (readable terminal logs by default):
agent.print_response("Give me a short summary.", max_rounds=4, stream=True, stream_events=True)
# Raw JSON lines:
agent.print_response("Give me a short summary.", max_rounds=4, stream=True, stream_events=True, event_format="json")

Autoresearch mode (persistent execution)

autoresearch=True enables persistent run behavior. The model is expected to end only when it calls agent.terminate(...), or when user/system limits stop the run.

agent = Agent.MTPAgent(
    provider=provider,
    tools=tools,
    autoresearch=True,
    research_instructions=(
        "Keep working until requirements are satisfied and verified. "
        "Call agent.terminate with reason+summary only when complete."
    ),
    debug_mode=True,
)

agent.print_response(
    "Compute the result and verify with tools. Terminate only after completion.",
    max_rounds=12,
    stream=True,
    stream_events=True,
)

Persist conversation sessions (JSON database)

from mtp import Agent, JsonSessionStore
from mtp.providers import OpenAI

session_store = JsonSessionStore(db_path="tmp/mtp_json_db")
agent = Agent.MTPAgent(provider=OpenAI(model="gpt-4o"), tools=tools, session_store=session_store)

agent.run("Remember this: project codename is Atlas.", session_id="chat-1", user_id="u1")
agent.run("What is the project codename?", session_id="chat-1", user_id="u1")

PostgreSQL and MySQL session stores are also available:

from mtp import PostgresSessionStore, MySQLSessionStore

pg_store = PostgresSessionStore(db_url="postgresql://user:pass@localhost:5432/mtp")
my_store = MySQLSessionStore(
    host="localhost",
    user="root",
    password="secret",
    database="mtp",
    port=3306,
)

Run examples

python examples/quickstart.py
python examples/groq_agent.py
python examples/groq_agent_events.py
python examples/ollama_agent.py
python examples/lmstudio_agent.py
python examples/mcp_stdio_server.py

Interactive TUI with Local Inference

# Install with local inference support
pip install -e ".[ollama,lmstudio]"

# Start TUI
mtp tui

# Switch to local provider
/backend ollama

# Follow interactive setup to select model
# Start chatting with your local LLM!

Streamlit UI

pip install -e ".[groq,dotenv,ui-streamlit]"
streamlit run examples/streamlit_groq_agent_chat.py

Agent OS

pip install -e ".[dotenv,ui-streamlit,groq,openai,openrouter]"
mtp agent-os

Docs map

• Quickstart
• Agent API Reference
• Storage and Sessions
• Providers
• Local Inference (LM Studio + Ollama)
• Creating Tools
• Events Contract
• Architecture
• Project Direction
• Protocol Spec
• Local Toolkits
• Groq Integration
• Transport
• MCP Interop Adapter
• Publishing
• Testing

Repository structure

• src/mtp/protocol.py: Core protocol entities (ToolSpec, ToolCall, ExecutionPlan, etc.).
• src/mtp/schema.py: Versioned envelope + execution plan validation.
• src/mtp/policy.py: Risk policy (allow / ask / deny).
• src/mtp/runtime.py: Tool registry, lazy loading, caching, batch execution.
• src/mtp/agent.py: Agent loop around provider + runtime.
• src/mtp/toolkits/: Local toolkits (calculator, file, python, shell).
• src/mtp/transport/: Envelope transport over stdio and HTTP.
• src/mtp/mcp.py: MCP-compatible JSON-RPC adapter around ToolRegistry.
• src/mtp/providers/: Provider adapters (MockPlannerProvider + OpenAI/LMStudio/Ollama/Groq/OpenRouter/Gemini/Anthropic/SambaNova/Cerebras/DeepSeek/Mistral/Cohere/TogetherAI/FireworksAI).
• docs/: documentation and implementation guides.

Contributors

Created by Prajwal Ghadge with contributions from Himesh Mehta.

See CONTRIBUTORS.md for the full list of contributors.

License

MIT License - see LICENSE file for details.