Goose AI and Other Terminal Agents: A Look Without the Subscription
What the open-source agent Goose can do, how it works with local models and MCP, and how it differs from desktop Doka and other CLI agents.
Goose often gets described as just another terminal chat, but that's not quite right. It doesn't just write an answer — it runs commands, changes files, and connects external tools through MCP. It belongs to the same category as other coding agents: it takes a task and works through several steps to a result.
What Goose AI is
goose is an open agent started at Block, available as a CLI and a desktop app. It isn't locked to one model: you can pick a cloud provider or a local model server. Its capabilities extend through MCP servers.
Search also turns up an old service called GooseAI with its own API — that's a
different product. If you're looking for the modern terminal agent, look for the
goose repository and docs specifically.
What's good about it
- Open source. The agent can be studied, built, and tuned for your own workflow.
- CLI and desktop. The terminal suits developers, and the app lowers the barrier to entry.
- Choice of model. Various providers are supported, including local ones.
- MCP extensions. You can connect GitHub, a browser, databases, and your own tools.
There's no required subscription for the agent itself. But whether the whole setup stays free depends on the model: a local one runs on your own hardware, a cloud API gets billed by the provider.
Where it gets complicated
A terminal agent gets serious permissions. A mistake in a command can change a lot of files, and a connected MCP server can send a message or update a task. Before working with it, you need to understand the permission model, set up a working folder, and avoid running the agent with administrator rights.
The second question is the model. A small local LLM saves money and keeps things private, but holds up worse over a long chain of actions. A strong cloud provider works more confidently, but code and context go outward.
"Works locally" refers to the model and tools you chose, not automatically to the whole install. Connect a cloud API, and requests will go to that provider.
Goose or Doka
Both agents work with files, the terminal, different models, and MCP. The difference is more in the starting experience.
Goose feels natural in a developer's CLI and appeals to people ready to assemble their own configuration out of a provider and extensions. Doka focuses on a simple desktop start: local models download from a catalog, projects and tools are available from the interface, and the terminal stays inside the agent's own process.
Better to choose not by a feature checklist but by one real task. Give both agents a small repository, the same model, and a constraint of "plan first, then make changes." Compare where it's clearer to confirm actions and check the result.
What other terminal agents exist
Alongside Goose, people usually consider Aider, OpenCode, Claude Code, and Qwen Code. For more general tasks there's Hermes Agent Desktop, and for an always-on gateway with messengers, OpenClaw. Each has its own focus: collaborative code editing, autonomous steps, choice of providers, or integration with a specific model family. Four questions matter more than the brand:
- Can you connect the model you need.
- How does the agent ask permission for risky actions.
- Is there MCP or another way to add tools.
- Is it convenient to see a diff and roll back a bad edit.
If you want to try the same class of task with no required subscription, download Doka and start with the built-in local model. How programs like this are built is covered separately in the article on AI in the terminal.