Local models are fast, free, and private. But ask one to call a tool and it falls apart. Wrong function names, broken parameters, infinite loops.

The models are capable. The framework wasn't.

OnsetLab makes 3B-7B models do reliable tool calling through a hybrid REWOO/ReAct architecture. The framework handles planning, execution, and error recovery. The model only does what it's good at: one step at a time.

Quick Start

pip install onsetlab

Requires Ollama running locally with a model pulled:

ollama pull phi3.5

from onsetlab import Agent
from onsetlab.tools import Calculator, DateTime

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

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

The agent routes the query, builds an execution plan, calls the right tool, and returns the answer. No prompt engineering required.

Architecture

The Router classifies queries as tool-needed or direct-answer using the model itself. The Planner generates structured THINK -> PLAN steps with auto-generated tool rules from JSON schemas. The Executor resolves dependencies and runs tools in order. If planning fails, the ReAct Fallback switches to iterative Thought -> Action -> Observation loops to recover.

Built-in Tools

More tools will be added over time.

MCP Servers

Connect any MCP-compatible server to give your agent access to external tools like GitHub, Slack, Notion, and more.

from onsetlab import Agent, MCPServer

server = MCPServer.from_registry("filesystem", extra_args=["/path/to/dir"])

agent = Agent("phi3.5")
agent.add_mcp_server(server)

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

agent.disconnect_mcp_servers()

Any MCP server available via npm works too. See the docs for examples.

Built-in registry: filesystem · github · slack · notion · google_calendar · tavily

CLI

python -m onsetlab                                          # interactive chat
python -m onsetlab --model qwen2.5:7b                       # specify model
python -m onsetlab benchmark --model phi3.5 --verbose        # validate a model
python -m onsetlab benchmark --compare phi3.5,qwen2.5:7b    # compare models
python -m onsetlab export --format docker -o ./my-agent      # export as Docker
python -m onsetlab export --format config -o agent.yaml      # export as YAML

Export formats: YAML (portable config), Docker (Dockerfile + compose + Ollama), vLLM (GPU-accelerated), Script (standalone .py file). See Export & Deploy docs for details.

Tested Models

Works with any Ollama model. Run python -m onsetlab benchmark --model your-model to verify.