Plugin & slash commands¶

One folder in the repository installs three things at once: the MCP server, a navigation skill that tells the agent which tool answers which question, and two slash commands. There is no MCP config file to edit — installing the plugin registers the server.

Install¶

Client	How
Claude Code	`/plugin marketplace add harshithsunku/mcp-gtags-server` then `/plugin install mcp-gtags-server@mcp-gtags-server`
Codex	`codex plugin marketplace add harshithsunku/mcp-gtags-server`, then install it from the Plugins Directory in the ChatGPT desktop app
Cursor	Cursor loads Agent Plugins 1.0.0 but installs from the Cursor Marketplace or a team marketplace; until then use the manual entry

Install one way, not both

A plugin plus a manual entry runs two servers, and every tool shows up twice under different names. Switching to the plugin? claude mcp remove gtags first.

Slash commands¶

These are MCP prompts, so they cost no tool-schema context at all. They work with any installation method in clients that support MCP prompts (Claude Code, Cursor).

`/gtags:impact [git_ref]`¶

Change-impact analysis, and the replacement for the old blast_radius tool:

run git diff <ref> (default HEAD) and list each changed C/C++ function;
call find_callers on each one;
report the callers ranked by risk, with path:line, naming those that likely need updating.

`/gtags:explain <symbol>`¶

Definition → body → callers → a short explanation covering purpose, inputs and outputs, side effects and locking, config-dependent variants, and the most important callers.

What the skill does¶

skills/c-code-navigation/SKILL.md is model-invoked: the agent loads it when a question looks like C/C++ navigation. It carries the routing table (which tool answers which question), how to read the compact output, the project_root rule, kernel specifics such as active_config, and the change-impact recipe. It also states plainly when to keep using grep — string literals, comments, Kconfig and Makefiles.

Layout¶

plugin/
├── plugin.json                      # Agent Plugins 1.0.0 (Codex, Cursor)
├── mcp.json                         # portable MCP server config
├── .claude-plugin/plugin.json       # Claude Code manifest
├── .mcp.json                        # Claude Code MCP server config
└── skills/c-code-navigation/SKILL.md

Marketplace manifests for all three clients live at the repository root (.claude-plugin/, .agents/plugins/, .cursor-plugin/) and point at that one folder.

Version pinning¶

Both MCP configs launch the server as:

{ "command": "uvx", "args": ["--from", "mcp-gtags-server>=2,<3", "mcp-gtags-server"] }

The plugin ships a skill and prompts that describe the 2.x tool surface, so it pins the server to 2.x: a future major that renames tools can't silently run under the old skill, while every 2.x fix still arrives automatically. Manual installs stay unpinned (uvx mcp-gtags-server).

Working directory¶

Agent Plugins clients launch the server from the plugin's folder, not your project. The plugin therefore sets GTAGS_MCP_CWD_FALLBACK=0, so the server asks for project_root instead of indexing the plugin directory. Claude Code also answers roots/list with your workspace, which resolves the repository automatically.