Goose — AI Developer Agent by Block

What it is

Goose is an open-source AI developer agent built by Block (formerly Square). It runs in your terminal, reads your codebase, executes shell commands, and edits files based on natural-language instructions. Unlike chatbot-style coding assistants, Goose operates autonomously within a session, maintaining context across multiple interactions.

The target audience is developers who want an AI pair programmer that lives in the terminal rather than inside an IDE. Goose supports extensible toolkits and connects to MCP servers for additional capabilities.

How it saves time or tokens

Goose's session memory means you do not need to re-explain project context on every interaction. It retains file edits, command outputs, and conversation history across a session. The toolkit architecture lets you preload domain-specific tools so the agent spends fewer tokens discovering how to interact with your stack.

By running commands directly in the terminal, Goose avoids the copy-paste loop between a chat UI and your editor. Multi-step tasks like 'create a new API endpoint, write tests, and run them' execute in a single session without manual intervention.

How to use

1. Install Goose via Homebrew or pipx:

brew install block/goose/goose
# or
pipx install goose-ai

1. Start a session and give instructions:

goose session start

1. Chat naturally. Goose reads your codebase, runs commands, and edits files. End the session when done:

# Inside the session, type your request:
> Add a health check endpoint to the FastAPI app and write a test for it

Example

A typical Goose session for scaffolding a Python CLI:

$ goose session start
Goose> Create a Python CLI tool called 'logparse' that reads
       nginx access logs, counts status codes, and outputs
       a summary table. Use click for argument parsing.

# Goose will:
# 1. Create the project structure
# 2. Write logparse/cli.py with click commands
# 3. Add requirements.txt
# 4. Run a test with sample data

Related on TokRepo

• AI agent tools -- Compare Goose with other AI coding agents
• Automation tools -- Terminal-based developer productivity tools

Common pitfalls

• Goose executes shell commands with your user permissions. Review destructive commands (rm, git push --force) before confirming.
• Session memory grows with conversation length. For large codebases, scope your instructions to specific directories to avoid context overflow.
• Toolkit configuration happens before session start. Adding a new toolkit mid-session requires restarting.
• Goose depends on an LLM API key (Claude, GPT, or local models). Ensure your API key is set in the environment or config before starting.