goose

Working on a code repository from the terminal

Install the CLI, run goose configure to choose a provider and model, then start goose session inside the repository. The Developer extension reads files, edits them with find-and-replace and runs commands such as tests and linters. Add a .goosehints file with build commands and conventions so they do not need repeating. Check which permission mode is active with /mode before asking for changes, and review the diff in git rather than relying on the chat summary.

Running a desktop assistant for non-code tasks

goose Desktop gives the same agent a chat interface with extension and settings screens, which suits research, document drafting, spreadsheet analysis or browser-based automation through Computer Controller. Enable only the extensions the task needs, since each one adds tools to the model context. For tasks that touch personal folders or accounts, choose Manual or Smart Approval and keep an eye on each tool call before it runs.

Using goose with private or local models

Select the Local provider in Settings, search Hugging Face for a supported GGUF model and download it, or point goose at Ollama, LM Studio, Docker Model Runner or an internal OpenAI-compatible endpoint. Pick a model with native tool calling where possible, because tool emulation is less reliable. Confirm that no cloud provider remains configured as a fallback and that the model fits in available GPU or system memory.

Automating recurring work with recipes

Capture a successful session as a recipe, add parameters for the parts that change and define a response schema if another system will read the output. Store recipes in the repository so changes are reviewed like code, and schedule them with a cron expression for reports or maintenance checks. Since 1.57.0 scheduled runs use the recipe settings, so verify the provider, model and extensions declared in the file.

Pull request review in GitHub Actions

The CI/CD tutorial installs goose in a workflow, configures a provider from repository secrets and runs a headless prompt against the pull request. Give the job the narrowest token permissions it needs and pin the goose version rather than following the stable channel. Treat the posted review as an input for a human reviewer, and test the workflow on a fork before enabling it on protected branches.

Rolling goose out to a team

Publish an extension allowlist over HTTPS and set GOOSE_ALLOWLIST on managed machines, decide on a default permission mode and document which providers are approved for company data. Teams that need a branded build can follow the custom distribution guide to preconfigure providers and bundled extensions. Use the usage data setting to confirm the telemetry choice, and keep logs available for incident review.