Terminal cockpit for AI agents and MCP server
unterm, developed by Zhitongblog, is a terminal and Model Context Protocol server that acts as a cockpit for AI agents. The app exposes structured terminal controls so language models can execute commands, monitor output, and interact with local files to run automation. It highlights programmatic command execution, observable change review, and visual terminal inspection. Target users include developers, AI researchers, and localization engineers seeking controlled agent-driven terminal workflows.
What tasks can you actually use it for?
The app targets terminal-centered automation, translating high-level agent instructions into concrete shell activity. It supports common project work such as software development tasks, system administration, and AI-driven text localization by letting models perform command runs, read command output, and manipulate project files. Practical outcomes include automated test runs, updating i18n files, and scripted environment changes, where an agent handles repetitive terminal steps instead of manual execution.
How controllable are the agent's actions?
Control is built into the workflow through multiple safety mechanisms that give humans a chance to inspect or reverse agent changes. The app provides:
- isolated worktrees for safe code manipulation,
- a review and rollback system for file diffs,
- a human inbox for verification before irreversible actions.
Those elements combine to reduce the risk of unintended system changes by requiring observable steps and review points before finalizing edits.
What file formats, methods, and integrations does it accept?
The app exposes an extensive MCP surface, including over 99 methods across multiple namespaces, so agents can perform granular terminal and filesystem operations. It runs as an MCP server and integrates with MCP-compliant clients such as Claude Desktop, Cursor, or Aider. The server is cross-platform and operates on macOS, Linux, and Windows, enabling agents to invoke commands and retrieve terminal state from those environments.
Is it suitable for non-technical users or team workflows?
The app is oriented toward users comfortable with agent workflows and terminal concepts rather than end users with no technical background. Its design favors delegation of scripted tasks to agents while keeping human oversight. Teams that need traceability and rollback for automated edits will find the observable controls useful, whereas individuals unfamiliar with MCP clients or agent orchestration face a learning curve to adopt it effectively.
Who should adopt unterm?
unterm is a practical choice for developers, researchers, and localization engineers who need programmatic, observable automation inside a terminal environment. Its emphasis on controlled agent actions and isolated change handling suits teams that accept an initial setup and oversight overhead. Expect a requirement for MCP tooling familiarity and review workflows before delegating high-stakes tasks to agents.
