Ollama Provider

01Install

Ollama runs open-source models locally. No cloud API key required. Supports tool calling, streaming, and thinking/reasoning traces.

Install

bashpip install "mtpx[ollama]"

Or install the SDK directly:

bashpip install ollama

02Setup

Setup

Install Ollama from ollama.com
Pull a model:

bash   ollama pull qwen3:1.7b

Verify the server is running:

bash   ollama list

No API key is required for local usage. For secured remote hosts, set OLLAMA_API_KEY in your .env file.

03Quick Start

Quick Start

pythonfrom mtp import Agent
from mtp.providers import Ollama

# No API key needed for local Ollama
provider = Ollama(model="qwen3:1.7b")
tools = Agent.ToolRegistry()
agent = Agent(provider=provider, tools=tools)

reply = agent.run_loop("What is 25 * 4 + 10?")
print(reply)

04Parameters

Parameters

Parameter	Type	Default	Description
`model`	`str`	`"qwen3"`	Ollama model name (must be pulled first)
`host`	`str \| None`	`None`	Ollama server URL (default: `http://localhost:11434`)
`api_key`	`str \| None`	`None`	Optional API key for secured hosts (falls back to `OLLAMA_API_KEY` env var)
`options`	`dict \| None`	`None`	Ollama-specific options (e.g., `{"temperature": 0, "num_ctx": 4096}`)
`format`	`dict \| str \| None`	`None`	Output format constraint (e.g., `"json"` or a JSON schema dict)
`keep_alive`	`float \| str \| None`	`None`	How long to keep model in memory (e.g., `"5m"`, `300`)
`think`	`bool \| None`	`None`	Enable thinking/reasoning traces (for models that support it)
`client`	`Any \| None`	`None`	Pre-configured `ollama.Client` instance

05Capabilities

Capabilities

Capability	Value
Tool calling	Yes
Parallel tool calls	Yes
Input modalities	text, image
Streaming	Yes (both `stream_next_action` and `finalize_stream`)
Usage metrics	Rich
Reasoning metadata	Yes (when `think=True`)
Native async	No (uses thread fallback)

06Recommended Models

Recommended Models

qwen3:1.7b — Lightweight, fast, good tool calling
qwen3:4b — Better quality, still fast
qwen3:8b — Strong tool calling
llama3.2:3b — Good general purpose
mistral:7b — Solid alternative
deepseek-r1:1.5b — Reasoning model with thinking traces

07Thinking/Reasoning

Thinking/Reasoning

Enable thinking traces for models that support it:

pythonprovider = Ollama(
    model="qwen3:1.7b",
    think=True,
    options={"temperature": 0},
)

When think=True, the provider:

Captures thinking tokens from the model
Surfaces them in action_meta["reasoning"]
Reports supports_reasoning_metadata=True in capabilities

08Streaming

Streaming

Ollama supports both planning and finalization streaming:

pythonfor event in agent.run_loop_events(
    "Calculate 25 * 4",
    max_rounds=2,
    stream_final=True,
):
    print(event)

09Full Example

Full Example

pythonfrom mtp import Agent
from mtp.providers import Ollama

provider = Ollama(
    model="qwen3:1.7b",
    host="http://localhost:11434",
    think=True,
    options={"temperature": 0},
)

tools = Agent.ToolRegistry()
agent = Agent(provider=provider, tools=tools, debug_mode=True)

reply = agent.run_loop(
    "Calculate (25 * 4) + 10",
    max_rounds=3,
)
print(reply)

10Troubleshooting

Troubleshooting

Model not found: Run ollama pull <model> first.
Connection refused: Ensure Ollama is running (ollama serve).
No tool calls: Use a model that supports tool calling (Qwen 3, Llama 3.2+).
Slow first request: Model loading into memory takes time on first call.

11Source

Source

src/mtp/providers/ollama_provider.py

Use this when

Read the system