MCP server list, the one on your disk

Two things share the phrase. First sense: your MCP client's list of servers it knows about, stored as a JSON file on disk at a client-specific path (Cursor uses ~/.cursor/mcp.json, Claude Desktop uses ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, Windsurf uses ~/.codeium/windsurf/mcp_config.json, and so on). Second sense: the list of tools a single MCP server exposes, returned over JSON-RPC 2.0 when the client sends a tools/list method. Both are literal lists. Neither is a marketing directory. If you are searching for 'mcp server list' hoping to see tools a specific server exposes, this page shows Terminator's 31-entry tool list below. If you are looking for where your own client stores its list, the 12-path map is also here.

From crates/terminator-mcp-agent/config.js lines 50 to 100: Cursor uses ~/.cursor/mcp.json. Claude Desktop uses ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows. Windsurf uses ~/.codeium/windsurf/mcp_config.json. Cline lives inside VS Code globalStorage under saoudrizwan.claude-dev/settings/cline_mcp_settings.json. RooCode is also under VS Code globalStorage at rooveterinaryinc.roo-cline/settings/mcp_settings.json. Witsy uses Application Support/Witsy/settings.json. Enconvo uses ~/.config/enconvo/mcp_config.json. BoltAI uses ~/.boltai/mcp.json. Amazon Bedrock uses ~/Amazon Bedrock Client/mcp_config.json. Amazon Q uses ~/.aws/amazonq/mcp.json. VS Code and VS Code Insiders are the two exceptions: there is no stable JSON path, so the installer shells out to code --add-mcp or code-insiders --add-mcp instead.

Because VS Code owns its MCP list internally and exposes a stable CLI flag instead. Look at line 82 of config.js: the vscode entry is { type: 'command', command: 'code' }. When you pick vscode in the --add-to-app wizard, the installer runs code --add-mcp <json-string> instead of touching a file. The vscode-insiders entry does the same with code-insiders. This is why config.js has two type kinds: file targets get their JSON merged in-place, command targets get an execFileSync call.

A flat JSON object with an mcpServers key. Every client listed above shares this shape. Example: { "mcpServers": { "terminator-mcp-agent": { "command": "npx", "args": ["-y", "terminator-mcp-agent@latest"] } } }. Each entry is a key (the server name your client will show in its UI) mapped to an object with command and args. The command is the executable the client spawns as a child process; it speaks MCP over that child's stdin and stdout. You can add as many entries as you like; the client reads the whole list at startup and spawns one subprocess per entry.

At compile time, by the build script. crates/terminator-mcp-agent/build.rs line 31 has a function extract_mcp_tools() that reads src/server.rs, scrolls until it finds the line containing 'let result = match tool_name', and then collects every string literal that matches the 'tool_name' => match arm pattern. It filters for lowercase-and-underscore names, deduplicates, joins with commas, and emits cargo:rustc-env=MCP_TOOLS=<list> so the value is baked into the binary. prompt.rs line 10 does let mcp_tools = env!("MCP_TOOLS") and pastes the list into the system instructions at line 99. The result: the tool list the LLM sees, the list the dispatcher matches against, and the list the build scanner extracts are always the same list, because they are all read from the same place in server.rs.

Thirty-one: get_window_tree, get_applications_and_windows_list, click_element, type_into_element, press_key, press_key_global, validate_element, wait_for_element, activate_element, navigate_browser, execute_browser_script, open_application, scroll_element, delay, run_command, mouse_drag, highlight_element, select_option, set_selected, invoke_element, set_value, execute_sequence, stop_highlighting, stop_execution, gemini_computer_use, read_file, write_file, edit_file, copy_content, glob_files, grep_files. Five groups: discovery (read the tree), action (click or type), browser bridge (DOM access via the Terminator Bridge extension), filesystem (typed file ops), orchestration (execute_sequence plus stop), and two escape hatches (run_command and gemini_computer_use).

Use the MCP Inspector or a JSON-RPC client directly. npx @modelcontextprotocol/inspector npx terminator-mcp-agent@latest starts a web UI that connects to the server over stdio and shows the live tools/list response. For raw verification, spawn the server yourself and send it a JSON-RPC frame: { "jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {} }. The response contains a tools array with name and description for every entry. This is authoritative in a way a docs page is not; you are reading the list the server itself produces at boot.

No. The JSON object key is the name; object keys cannot repeat. If you add terminator-mcp-agent twice, the second overwrites the first. This is also why the installer's writeConfigFile function reads the existing config, merges a new entry into mcpServers, and writes the result back: it is a dictionary merge, not an append. If you want two instances of the same server (for example, one on stdio and one on HTTP), give them different keys.

Functionally yes, protocol-wise. If the server speaks Model Context Protocol correctly over its declared transport (stdio or HTTP), every compliant client can list its tools and dispatch calls. Ergonomically the clients differ. Claude Desktop and Cursor render the tool list as a sidebar. Claude Code shows it via claude mcp list. VS Code surfaces MCP servers in its extensions model. Cline and RooCode expose the list in their respective panels. The underlying list of servers in the JSON file is identical in shape across clients; the client's UI is what changes.

Because the 12-path map is what users would otherwise have to memorize. Finding Claude Desktop's config on Windows at %APPDATA%\Claude\claude_desktop_config.json is not the same keystroke as finding it on macOS at ~/Library/Application Support/Claude/claude_desktop_config.json, and neither matches the VS Code --add-mcp flow. config.js in the agent's npm package encodes all 12 of those resolutions so one command, --add-to-app, writes to the correct file or runs the correct CLI. Users do not have to learn the map; the installer already knows it. That file is the anchor of this page.