volt-llm-client 0.2.0

A lightweight client for local LLM APIs like Ollama or OpenWebUI

volt-llm-client

volt-llm-client is a lightweight Python client for interacting with local LLM APIs such as Ollama or OpenWebUI.
It supports prompt completion, multi-turn conversations, and model listing — perfect for scripts, prototypes, or CLI tools.

---

Features

• Send single prompts or full conversations
• Compatible with Ollama/OpenWebUI-style APIs
• Optional MCP tool calling via a Docker Desktop MCP Toolkit gateway
• Token-based authentication support
• Clean log output using volt-logger
• Minimal dependencies (requests only; fastmcp for the optional mcp extra)

---

Installation

Install both volt-llm-client and its logging dependency:

pip install volt-llm-client

For local development:

git clone https://github.com/stuarttempleton/volt-llm-client.git
cd volt-llm-client
pip install -r requirements.txt

---

Usage Example

import os
from voltllmclient import LLMClient

llm = LLMClient(
    token=os.getenv("LLM_API_TOKEN"),
    model="Gemma4"
)

reply = llm.send_prompt("What is the capital of France?")
print(reply)

You can also send a full conversation:

messages = [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "Who won the World Cup in 2018?" }
]

response = llm.send_conversation(messages)
print(response)

---

Maintaining Context with LLMConversation

from voltllmclient import LLMConversation
import os

conv = LLMConversation(model="gemma4", token=os.getenv("LLM_API_TOKEN"))

response = conv.send("What is quantum computing?")
print(response)

response = conv.send_with_full_context("How is it different from classical computing?")
print(response)

conv.save_transcript("session.json")

---

MCP Tool Calling

Lets the model call tools from your MCP servers. Requires the mcp extra:

pip install volt-llm-client[mcp]

from voltllmclient import LLMClient, MCPToolProvider

# Talks to a Docker Desktop MCP Toolkit gateway. Pass the name of the profile holding
# the servers you want — without one, the gateway only exposes its own meta-tools.
provider = MCPToolProvider(profile="my_profile")

if provider.connect():
    llm = LLMClient(model="gemma4:26b", mcp=provider)
    print(llm.send_with_tools([{"role": "user", "content": "What is in my log files?"}]))

provider.close()

LLMConversation can do the plumbing for you — hand it a profile name and it builds, connects, and closes the provider itself:

from voltllmclient import LLMConversation

with LLMConversation(model="gemma4", mcp="my_profile") as conv:
    print(conv.send("What is in my log files?"))

Pass a MCPToolProvider instead of a string when you want the filtering options, or to share one gateway across several conversations. A provider you built is yours: close() leaves it running.

List your profiles with docker mcp profile list, and check what a profile exposes with docker mcp tools ls --gateway-arg=--profile=<name>.

A profile is required. Without one the gateway advertises only its own meta-tools (mcp-add, mcp-config-set, ...) which can rewrite your MCP setup, so connect() refuses and returns False rather than handing those to a model. Pass args= instead to drive a non-Docker stdio MCP server:

MCPToolProvider(command="my-mcp-server", args=["serve", "--stdio"])

Config is per instance — nothing is read from the environment — so several providers with different profiles and filters can run side by side in one process.

send_with_tools loops until the model stops requesting tools (capped by max_tool_rounds=5). If the gateway is unreachable, connect() logs a warning and returns False — pass mcp=None and the prompt is still answered without tools.

The gateway's own stderr is captured to a log file rather than the console (it is noisy on success and prints a stack trace when Docker Desktop is down). On failure the warning quotes the relevant line from it and tells you where the full output is.

Each provider gets its own temp log (<tempdir>/volt-mcp-gateway-<pid>-<random>.log) so concurrent instances never clobber each other. It is deleted on close() unless the connection failed, in which case it is kept for you to read. Pass MCPToolProvider(log_file="path/to/gateway.log") to choose the location yourself — a log you supply is never deleted.

Narrowing the tool list

Every advertised tool costs prompt tokens on each round, and a gateway with dozens of similarly-named tools makes a small model more likely to pick the wrong one. Filter client-side:

MCPToolProvider(include="get_*,search_*")        # glob, comma-separated
MCPToolProvider(tools=["get_file", "list_dirs"]) # exact names

CLI

python -m voltllmclient.client gemma4 "What is the capital of France?"

# with tools from an MCP profile, narrowed to the ones the model needs
python -m voltllmclient.client gemma4 "What is in my log files?" \
    --profile my_profile --tools "get_*,search_*"

Option	Default	Meaning
--url	http://localhost:11434	API base url (domain:port, not an endpoint)
--profile	none	Docker MCP profile to expose; omit and no MCP is used at all
--tools	all	tool filter glob, comma separated
--timeout	120	request timeout in seconds; raise it if prompt processing is slow

---

Environment Variables

Set your API token via environment variable:

Unix/macOS:

export LLM_API_TOKEN=your_token_here

PowerShell:

$env:LLM_API_TOKEN = "your_token_here"

---

License

MIT