Skip to content
ochat
Search documentation

Use quotes for an exact phrase.

Search by topic, command, or code identifier.

    GitHub ↗

    Shell access

    Understand declared shell capabilities, host authorization, and runtime policy.

    View Markdown source ↗

    Shell access lets an agent use command-line programs: inspect a repository, run a test, or interact with a configured process. It is powerful enough to deserve an explicit design, not just an instruction telling the model to be careful.

    Ochat’s shell declarations describe runtimes and tools in ChatMD. The host decides which authority it will admit, and approval policy determines when an operation needs another decision. Begin with a narrow command and expand only when the task requires more access.

    Follow the shell-agent walkthrough. It uses a limited pwd tool and explains how to run it with the appropriate authorization. The declaration examples cover more patterns, but are not all standalone prompts or universal policies.

    Local and daemon hosts do not use interchangeable authorization switches. Read host modes and authorization before copying a command from one execution mode to another.

    PieceQuestion it answersGuide
    Runtime declarationWhere and under what execution settings do commands run?Runtime reference
    Tool declarationWhat command interface does the agent see?Shell tools
    Authority and confinementWhat resources and effects are actually permitted?Security guide
    Host authorizationHow does this local runner or daemon admit the declared access?Host integration
    Review and approvalWhich requests need a decision, and who or what makes it?Extensions and reviewers
    Durable recordsWhat is retained, audited, or interrupted across restarts?Persistence and audit

    The workspace supplies a location, not a security boundary by itself. Review the declared filesystem and network access, the actual confinement backend, and approval policy together. Unattended agents need a deliberate noninteractive policy; removing prompts for human approval does not reduce the consequences of the commands they can run.

    Return to tools or the documentation home.