goose 1.51.0 shipped on September 17, 2026. The release adds GPT-live API support and better connections to external ACP backends, while fixing MCP retries and several permission and session-storage issues. The execution boundary matters more than the feature count here. A coding agent that can edit files and run shell commands needs clear permissions as much as it needs a capable model.
If you need an open-source agent that can modify and test code, goose is one practical option. It comes as a desktop app, CLI, and API. It connects to cloud models from OpenAI, Anthropic, and Google, as well as local models served through Ollama or LM Studio.
The short version
- goose is open source under Apache 2.0.
- Desktop and CLI builds are available for macOS, Linux, and Windows.
- It supports more than 15 model providers and more than 70 MCP extensions.
- goose is free, but cloud inference is billed by the provider you choose.
- Local models need tool-calling support and enough context for agent work.
- Because goose can access the shell and filesystem, start with approval required for writes and commands.
What changed in 1.51
Version 1.51.0 combines new integrations with tighter execution behavior. It adds GPT-live API support, improves connections to external ACP backends, introduces an EUrouter provider, and can optionally ring the terminal bell when a turn completes or needs approval.
The release also fixes MCP preferred-version negotiation and HTTP retries. Interrupted headless runs now return failure, which matters in automation because a pipeline should not continue after an incomplete agent run.
Several fixes reduce obvious security risks: gateway pairing codes receive additional protection, session-storage permissions are restricted, scheduled-session content is filtered, and MCP app tools are bound more closely to their extension owners. These changes do not isolate the agent process by themselves.
Install and start a session
The official macOS package is available through Homebrew. A package manager is easier to inspect and update than piping a remote installer directly into a shell.
brew install block-goose-cli
goose configureLinux and Windows binaries are available from the release page. For CI, the project recommends pinning GOOSE_VERSION instead of depending on the moving stable tag.
After choosing a provider, start a session inside the repository you want goose to work on.
cd my-project
goose sessionKeep the first task narrow. A prompt such as "Find the cause of one failing test and show the proposed changes before editing" sets both scope and an approval point.
Cloud providers or a local model
goose supports OpenAI, Anthropic, Gemini, GitHub Copilot, OpenRouter, and many other providers. Existing subscriptions can also be used through ACP integrations for ChatGPT Codex, Claude Code CLI, and Gemini CLI. Charges come from the selected model service or subscription, not from goose itself.
For code that should not be sent to a hosted API, goose can connect to Ollama, LM Studio, or Docker Model Runner. Check tool-calling support before model size or benchmark scores. A model without tool calling can only provide chat completions and cannot use goose extensions effectively.
Context length is another practical limit. The documentation suggests increasing OLLAMA_CONTEXT_LENGTH when an Ollama model starts ignoring extensions or .goosehints. More context also consumes more memory, so model size and quantization still matter on local hardware.
Set narrow permissions first
goose provides three tool-permission levels: Always Allow, Ask Before, and Never Allow. A sensible starting point is to allow read-only operations, require approval for file edits and system commands, and block credential access or deletion unless a task genuinely needs it.
The project’s security guidance explicitly warns about prompt injection when goose reads internet content or other untrusted data. Use a dedicated VM or a container with limited privileges for sensitive repositories, and connect only MCP extensions you have reviewed. Tool permissions reduce the blast radius, but they are not an operating-system sandbox.
Who should use it
goose fits developers who want to switch model providers without replacing the whole agent workflow, combine MCP extensions, or compare local and hosted models from one client. Offering both a desktop app and CLI also makes it more approachable than terminal-only agents.
It is a weaker fit if you expect a fully isolated runtime with no setup, or if you plan to rely on a small local model for dependable tool calls. Aider is a useful comparison when concise, terminal-first diff editing is the priority. OpenCode offers another open-source terminal agent experience. Claude Code is a managed alternative with tighter model integration, but it is not open source.
Tested scope and remaining limits
The Linux x86_64 musl binary for v1.51.0 returned 1.51.0 from goose --version. Its help output exposed the configure, doctor, session, run, review, schedule, and gateway commands. A live model call was outside this check because it requires provider credentials.
Version 1.51.0 fixes several permission and failure-handling problems, but it cannot guarantee that agent-generated commands are safe. Start in a test repository with limited privileges, then expand automatic approvals only after the workflow is predictable.
Tested and written: 2026-09-20
Sources
- goose v1.51.0 release notes: https://github.com/aaif-goose/goose/releases/tag/v1.51.0
- goose repository and Apache 2.0 license: https://github.com/aaif-goose/goose
- goose installation guide: https://goose-docs.ai/docs/getting-started/installation
- goose provider guide: https://goose-docs.ai/docs/getting-started/providers
- goose tool-permission guide: https://goose-docs.ai/docs/guides/managing-tools/tool-permissions
- goose security guidance: https://github.com/aaif-goose/goose/blob/main/SECURITY.md
- Aider repository: https://github.com/Aider-AI/aider
- OpenCode repository: https://github.com/anomalyco/opencode
- Claude Code overview: https://docs.anthropic.com/en/docs/claude-code/overview
Leave a Reply