Built-in MCP (Model Context Protocol) server for AI agent integration.

# Start MCP server over stdio
officecli mcp

# Register as MCP server in an AI client
officecli mcp <target>

# Unregister from an AI client
officecli mcp uninstall <target>

# Show registration status
officecli mcp list

OfficeCLI includes a built-in MCP server that exposes all document operations as tools over the Model Context Protocol (JSON-RPC 2.0, protocol version 2024-11-05). This allows AI agents (Claude, Cursor, VS Code Copilot, LM Studio, etc.) to directly call OfficeCLI commands.

# Register in Claude Code
officecli mcp claude

# Register in Cursor
officecli mcp cursor

# Check status across all clients
officecli mcp list

# Unregister from Claude Code
officecli mcp uninstall claude

The MCP server exposes a single tool named officecli with exactly one parameter, command — a raw CLI command line (string, or a pre-split argv array for values containing spaces/quotes), passed through verbatim to the same CLI engine as the terminal binary. There is no structured per-verb parameter schema (no file/path/props/parent fields) — the tool description explicitly says "Pass an officecli command line in command... it runs through the same CLI you'd use in a terminal."

{ "command": "add deck.pptx /slide[1] --type shape --prop text=Hi" }
{ "command": "set report.docx /body/p[1] --prop bold=true" }
{ "command": "view deck.pptx screenshot --page 2" }
{ "command": "help pptx shape" }
{ "command": "load_skill pptx" }

All CLI verbs are reachable this way: create, view (modes text/annotated/outline/stats/issues/html/svg/screenshot/forms), get, query, set, add, remove, move, swap, validate, batch, raw, raw-set, help, load_skill, and the rest of the CLI's command list — load_skill is a verb passed inside command, not a separate top-level MCP tool. Add --json to get/query/validate/view issues for structured output, same as the CLI.

v1.0.120+: the MCP tool is a thin shell over the officecli command line — the router in-process shells out to the CLI, so MCP and CLI behavior stay in lockstep by construction. The tool description carries an always-on delivery gate: before reporting a document finished, validate (schema), view issues + a view text placeholder scan (content), and a view screenshot visual audit (layout) — "validate passing is NOT delivery."

Invoked as {"command": "load_skill <name>"}, mirroring the CLI officecli load_skill <name>. Returns the embedded SKILL.md content for the named specialized skill (e.g. pptx, docx, xlsx, morph-ppt, pitch-deck, academic-paper, data-dashboard, financial-model) so agents can dynamically load just the scene-layer guide for the artifact they're producing without reinstalling anything.

Returns a strategy guide and format-specific property reference for the specified document format:

{ "command": "help", "format": "xlsx" }
{ "command": "help", "format": "pptx" }
{ "command": "help", "format": "docx" }

• The MCP server communicates over stdio (stdin/stdout) using JSON-RPC 2.0.
• All tool outputs use the structured JSON envelope format ({ "success": true/false, "data": ... }).
• The server runs as a single process — one instance per AI client session.

• Command Reference - All commands
• skills - Install skill definitions for agent clients
• Agent Guide - Wiki navigation for AI agents
• batch - Batch multiple operations

Based on OfficeCLI v1.0.127