onsetlab 0.1.0

Build, benchmark, and package local SLM agents. No API keys, no cloud.

OnsetLab

Make your local SLMs do actual work.

What

Agent framework for small language models (3B parameters) that achieves GPT-4 level tool use on your laptop. Uses REWOO architecture (plan → execute → verify) with ReAct fallback for robustness.

• No API keys - runs entirely on Ollama
• No fine-tuning - works out of the box with any Ollama model
• MCP support - connect to any MCP-compatible server
• Package & Deploy - export as Docker, config, or standalone script

Install

pip install onsetlab

Requires Ollama running locally.

Quick Start

from onsetlab import Agent

agent = Agent("phi3.5")
result = agent.run("What time is it in Tokyo?")
print(result.answer)

Adding Tools

Choose only the tools you need:

from onsetlab import Agent
from onsetlab.tools import Calculator, DateTime, UnitConverter, TextProcessor, RandomGenerator

agent = Agent(
    model="phi3.5",
    tools=[Calculator(), DateTime()]
)

result = agent.run("What's 15% tip on $84.50?")
print(result.answer)  # "The 15% tip on $84.50 is $12.68"

Available built-in tools:

• Calculator - math operations
• DateTime - current time, timezones, date math
• UnitConverter - convert between units
• TextProcessor - word count, search, transform text
• RandomGenerator - random numbers, strings, choices

MCP Servers

Connect to external services via Model Context Protocol:

from onsetlab import Agent, MCPServer

agent = Agent("phi3.5")

# Add filesystem access
agent.add_mcp_server(MCPServer.from_registry("filesystem", path="/my/project"))

# Add GitHub integration
agent.add_mcp_server(MCPServer.from_registry("github", token="ghp_..."))

result = agent.run("List all Python files in the project")
print(result.answer)

Registry servers: filesystem, github, slack, notion, google_calendar, tavily

Or configure any MCP server manually:

server = MCPServer(
    name="my-server",
    command="npx",
    args=["-y", "@some/mcp-server"],
    env={"API_KEY": "..."}
)
agent.add_mcp_server(server)

Packaging & Deployment

Export your agent for deployment:

# Export as Docker (includes docker-compose with Ollama)
agent.export("docker", "./my_agent/")

# Export as config file
agent.export("config", "my_agent.yaml")

# Export as standalone script
agent.export("binary", "my_agent.py")

Or via CLI:

python -m onsetlab export --format docker --output ./deployment/
cd deployment && docker-compose up --build

See docs/PACKAGING.md for full deployment guide.

Benchmarking

Compare model performance on tool-calling:

# Benchmark single model
python -m onsetlab benchmark --model phi3.5

# Compare models
python -m onsetlab benchmark --compare "phi3.5,qwen2.5:3b,llama3.2:3b"

See specs/BENCHMARKING.md for methodology.

Architecture

Task → Router → Strategy Selection
                    ↓
         ┌─────────┼─────────┐
         ↓         ↓         ↓
      DIRECT    REWOO     REACT
         ↓         ↓         ↓
                   └─────────┘
                       ↓
                   Verifier → Solver → Answer

1. Router - selects optimal strategy (DIRECT/REWOO/REACT)
2. Planner - creates step-by-step execution plan (REWOO)
3. Verifier - validates plan before execution
4. Executor - runs tools with dependency resolution
5. Solver - synthesizes final answer from results
6. ReAct fallback - recovers from planning failures

Options

agent = Agent(
    model="phi3.5",           # any Ollama model
    tools=[...],              # built-in tools
    mcp_servers=[...],        # MCP servers
    memory=True,              # conversation memory
    verify=True,              # pre-execution verification
    routing=True,             # intelligent strategy selection
    react_fallback=True,      # fallback on REWOO failure
    debug=False               # verbose logging
)

License

Apache 2.0