LLM Goose 🪿

LLM-powered testing, traces, and live chat for agent workflows

---

Goose is a Python library, CLI, and web dashboard for testing and debugging LLM agents.
Scaffold a gooseapp/, point Goose at your real query function, write goose.case(...) tests, then expose tools and live chat when you want the full dashboard loop.

Visit the landing page

What it looks like

Why Goose?

• Natural-language expectations – Describe the behavior you want and let Goose validate it.
• Tool call assertions – Check what your agent actually did, not just what it said.
• Full execution traces – Inspect messages, tool calls, tool outputs, and validation results.
• Live chat for iteration – Try agents in the dashboard while you develop.
• Hot reload – Re-run against updated code without restarting the app.

Choose your path

• Framework-agnostic quickstart – integrate Goose into an existing Python app, regardless of framework: docs/getting-started.md
• LangChain / LangGraph integration – keep your existing LangChain-style agent and add Goose around it: docs/integrations/langchain.md

If you are starting from zero, the framework-agnostic quickstart is the default path.

Core workflow

1. Scaffold the app with goose init
2. Point Goose at query(...) -> AgentResponse in gooseapp/conftest.py
3. Write cases in gooseapp/tests/ with goose.case(...)
4. Run the first loop with goose test list and goose test run
5. Expand into tools, chat, and hot reload through gooseapp/app.py

Install

Required for the first test run:

pip install llm-goose

Optional for the browser UI:

npm install -g @llm-goose/dashboard-cli

goose init creates the path

goose init

gooseapp/
├── README.md
├── __init__.py
├── app.py
├── conftest.py
└── tests/
    ├── __init__.py
    └── test_example.py

• app.py configures tools, live chat agents, and hot reload
• conftest.py wires the Goose fixture to your real query function
• tests/ holds the cases you run from the CLI or dashboard

See docs/goose-init.md for the full scaffold contract.

Minimal first test

from goose.testing import Goose


def test_agent_responds(goose: Goose) -> None:
    goose.case(
        query="Hello, what can you help me with?",
        expectations=[
            "Agent responds with a greeting or acknowledgment",
            "Agent describes its capabilities or offers assistance",
        ],
    )

goose is injected from the fixture you register in gooseapp/conftest.py. The full query -> fixture -> test path is documented in docs/getting-started.md.

Key commands

goose init                  # scaffold gooseapp/
goose test list gooseapp.tests
goose test run gooseapp.tests
goose api
goose-dashboard

Where to go next

• Framework-agnostic quickstart: docs/getting-started.md
• LangChain / LangGraph integration: docs/integrations/langchain.md
• Scaffold details: docs/goose-init.md
• Writing tests: docs/testing.md
• Running the API and CLI loop: docs/running-goose.md
• Using the dashboard: docs/dashboard.md

License

MIT License – see LICENSE for full text.