An LLM agent can drive zshot in several ways. Pick the one that fits how the agent runs.

MCP over HTTP

Run zshot as a server and connect the agent to the /mcp endpoint. Use this when the agent and zshot run as separate services, or when one zshot instance serves many agents. The endpoint is on unless --server-disable-mcp-endpoint is set. See MCP for the connection command.

MCP over stdio

Run zshot with --mcp-local-server and let the agent launch it as a local subprocess, talking over stdin and stdout. Use this for a desktop agent that manages its own tools, such as Claude Code or Claude Desktop.

CLI with the Agent Skill

Install the bundled Agent Skill with --install-skill and let the agent call the zshot binary directly. The skill teaches the agent what zshot does and how to invoke it correctly, with no server in between. Use this when the agent already runs shell commands.

In-tool guidance

zshot --agent-help prints a start-here guide written for an agent discovering the tool. From there, zshot --help <category> lists the flags in one category and zshot --help all lists every flag. The full documentation is also published for machines at llms.txt and llms-full.txt.