grade

Run the polygraph litmus on $ARGUMENTS using the run_litmus tool from the polygraph-litmus MCP server (bundled with this plugin). Report the letter grade, the one-line summary, and any failed category with its findings. If the grade is capped at B because Docker was unavailable, say so plainly.

run_litmus grades MCP servers, not skills. If $ARGUMENTS is a skill (a skill directory, or a request to grade or validate a skill), use the skill path instead — the run_skill_litmus tool or the grade-skill prompt — not run_litmus.

Token-gated servers

If the target is an https:// URL and run_litmus fails to connect because the server is token-gated (an auth error such as 401, invalid_token, or "No authorization provided"), litmus needs a bearer token — it connects as a fresh, independent client, so it has none of the user's existing session.

Don't immediately ask the user to paste one. First look for the token they already have: the same MCP client config that makes this server work in their agent. Only when that fails should you ask.

1. If the user already supplied a token (in $ARGUMENTS or the conversation), use it directly — skip discovery.
2. Otherwise, check the config files that exist and find the entry whose URL matches the target (ignore a trailing-slash difference). Common locations: project ./.mcp.json, ./.cursor/mcp.json, ./.vscode/mcp.json; user-level ~/.claude.json, the Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, the equivalent for the platform), and ~/.cursor/mcp.json. Read that entry's headers, resolving any ${VAR} placeholder from the environment.
3. If found, confirm before using it: name the source file, never print the token value, and state plainly that grading will make live, authenticated tool calls to the server as the user (read-only tools only). On confirmation, re-run run_litmus passing the token as bearer (strip a leading Bearer ), or as a header "Key: Value" for a non-Authorization scheme.
4. If nothing is found, the server may use OAuth. Tell the user a browser window will open for them to log in, then re-run run_litmus with interactive_auth: true — litmus opens the browser, captures the token via a local 127.0.0.1 callback, and uses it for this run only (never stored). If the user declines, or the server isn't OAuth, ask for a bearer token and say why: litmus connects as a fresh client and needs the same token their agent already uses for this server.

Never pass unsafe_host_exec, and never enable state-changing calls just to get past auth.

If run_litmus isn't in your available tools, it may just need loading (some clients surface MCP tools lazily) — try to load it before concluding it's missing. Only if the polygraph-litmus server genuinely isn't connected, tell me to enable the plugin (/plugin install polygraph@polygraphso, adding the marketplace first with /plugin marketplace add polygraphso/litmus if needed) and restart the session so the tools load. Do not suggest claude mcp add — the plugin already provides the server.