MCP Rollcall

mcp-rollcall

Every agent starts its configured MCP servers at session start and, when one fails, says failed to connect or Connection closed and nothing else. The server's stderr, the exit code, the missing directory, the port nobody is listening on - all of that is discarded, in Claude Code, Codex, Cursor, Gemini CLI, Copilot, VS Code, Windsurf and OpenCode alike. Each also keeps its servers in a different file. This skill reads every place a server can be configured for the host you are in, checks each one the way the host would try to start it, and on request actually starts it, does the MCP handshake and reports what came back.

The script is scripts/mcprollcall.py next to this file - run it from wherever this skill was installed (~/.claude/skills/mcp-rollcall/, ~/.agents/skills/mcp-rollcall/, ~/.cursor/skills/..., or a project's .agents/skills/). Stdlib-only Python 3.10+. It never writes a config file. Only --probe starts servers; everything else is static.

Which host

The script detects the host from the environment (CLAUDECODE, CODEX_SANDBOX, CURSOR_AGENT, GEMINI_CLI) and reads that host's files. If nothing identifies the host it reads every host's files and labels each row <host> <scope>. Pass --agent when you know better: --agent codex, --agent cursor,vscode, --agent all.

A name that appears in two hosts' files is two servers, not a clash; shadowed only means the same name twice inside one host. Claude Code's allow / deny lists apply to Claude Code's servers only, Gemini's mcp.allowed to Gemini's.

Pick the mode from what the user asked

--project DIR when they mean another project. --timeout N for servers that take long to start (first npx -y / uvx runs download). --json for processing.

Steps

1. Run the static roll call first. It is fast and touches nothing. For each server the status column is ok, warn, ERROR, shadowed, pending, disabled or DENIED; the first finding is printed beside it and the full list below. The static checks catch most failures on their own:

   • command-missing - the command is not on PATH or the path does not exist. The fix line says what to install or which full path to use.
   • windows-shim - a bare npx / uvx on Windows; a host that spawns without a shell (Claude Code and Cursor document this) needs cmd /c npx ... or the .cmd path. For other hosts it is a note, not a warning.
   • arg-path-missing / cwd-missing - an absolute path in args (a script, a directory to serve) or a Codex / Gemini cwd that is not there.
   • var-unset - ${VAR}, ${env:VAR}, a Codex bearer_token_env_var or env_http_headers variable, or a Gemini $VAR, with no value and no default.
   • env-empty / env-placeholder / header-placeholder - a token that was never filled in.
   • port-closed - the server (or an mcp-remote bridge in its args) points at localhost:PORT and nothing is listening. This is the usual reason for "Connection closed": the desktop app that hosts the server is not running.
   • shadowed - the same name in two scopes of one host; local beats project beats user.
   • approval - a Claude Code .mcp.json server the user has not approved yet.
   • denied / not-allowed / excluded - blocked by Claude Code's deniedMcpServers / allowedMcpServers or Gemini's mcp.excluded / mcp.allowed.
   • disabled - enabled = false (Codex, OpenCode) or disabledMcpjsonServers (Claude Code); the host skips it on purpose.
   • envfile-missing (Cursor), input-undefined (VS Code), short-timeout and required (Codex): host-specific facts printed as the host would act on them.

2. If the static pass is clean for the server they asked about, run --probe <name> --full. It spawns the server with its configured env and cwd, sends initialize, then tools/list, then terminates it. Read the result:

   • FAILED, exited before answering initialize - the stderr tail is printed; the last non-empty line is usually the exception. Quote it. That is the answer the host withheld.
   • FAILED, no initialize response within Ns - it started but never spoke MCP. Wrong command (a CLI that is not a server), wrong transport (an HTTP server configured as stdio), or a first-run download still going: rerun with a longer --timeout before concluding.
   • ok with start, tools and ~tokens - it works. slow-start (over 3 s) delays every session; suggest pre-installing the package instead of npx -y / uvx. heavy is thousands of tokens of tool definitions loaded per session; that is a cost the user may want to know about, not a fault.

3. Answer the question first. If they asked about one server, lead with its line and its finding, then the fix. Do not paste the whole table for twenty servers when one matters; give the count and the exceptions. Under --agent all, say which host each finding belongs to - the same package configured for two hosts fails or succeeds independently.

4. Probing has side effects the host also has: a server that talks to a desktop application (Blender, Photoshop, SolidWorks) will try to reach it, log that it could not, and still handshake. That is normal and not a failure. A server that refuses to start without its application is the one to report.

5. The claude.ai connectors (plugin:*) are not files; the tool lists the ones Claude Code has cached as needing authorization. The fix is in claude.ai's connector settings, not on disk, and you cannot do it from here - say so. Codex and Cursor OAuth servers get a TCP reachability check only.

6. Fixing is by hand, in the host's own file. Claude Code: claude mcp remove X / claude mcp add X ... rewrites the entry cleanly; editing ~/.claude.json directly while a session is open risks the app overwriting it. Codex: edit config.toml and restart. Cursor and VS Code: edit mcp.json; the editor reloads it. Gemini: gemini mcp add or edit settings.json. After the change, reconnect (/mcp in Claude Code, restart for Codex) and rerun the roll call to confirm.

Related, not overlapping

claude mcp list, codex mcp list and gemini mcp list print configured or connected / failed per server with no reason. /mcp in a Claude Code session reconnects and shows auth state. None shows stderr, checks paths and env before starting, or counts tools. skill-rollcall and settings-effective do the same job for skills and settings.

What this cannot do

It does not know a server's own conventions: a directory list a server splits on commas while the config uses semicolons looks fine statically and only the probe catches it. It does not authenticate: an HTTP server behind OAuth gets a TCP check, not a handshake. Host detection is best effort - a host that sets no environment marker is reported as all; pass --agent. It reads config files, not plugin bundles: Claude Code plugin servers are found by scanning the plugin cache for .mcp.json (best-effort), and Codex / Cursor / VS Code marketplace plugins are not read at all. Codex's .codex/config.toml is listed even when the project is not trusted, because trust state is not on disk. And a probe that succeeds here can still fail in the host if the host's PATH or environment differs from the shell you ran this from; the report prints the command it ran so the difference can be found.