Connect Claude over MCP

The Claude Code plugin gives Claude 14 MCP tools over your session library and three skills: one writes a report on a session or campaign, one finds and ranks the moments where testers got stuck, and one turns an open note into a code fix in your game's repo (Apply notes with an agent). Claude Code and claude.ai can also reach your Mulno Cloud sessions through the platform's own server (section 4); ChatGPT and Cursor are untried.

1. Load the plugin

The plugin is the plugin/ folder we send with a release. Studio installs mulno.exe beside itself, and the plugin finds it at %LOCALAPPDATA%\Mulno Studio\mulno.exe, so there is nothing to configure.

# Studio installed? The plugin finds mulno.exe itself.
claude --plugin-dir .\plugin

Or copy the folder to %USERPROFILE%\.claude\skills\mulno-playtest\ to load it in every session. Without Studio, point MULNO_EXE at a mulno.exe, and MULNO_MCP_ROOT at a library folder if its sessions aren't indexed yet.

2. Ask

Try "find moments where players got stuck in the latest session". The tools see whatever Studio, or mulno index, has already indexed. Every claim in a report cites a timestamp and a link.

3. Read-only unless you say otherwise

A tester controls every byte of the session files an agent reads, so an untrusted session could put instructions in front of an agent that holds write tools. The plugin therefore starts mulno mcp read-only: nine tools only read (list_sessions, get_session, get_events, get_timeline, across_sessions, get_transcript, search, get_frames, get_moment).

Set MULNO_MCP_WRITE=1 before starting Claude Code to turn on the five tools that write AI notes back onto a session's timeline (add_note, update_note, delete_note, reply, set_note_status). They can only ever add, edit or delete the agent's own notes, never a reviewer's and never media.

4. Remote MCP

The platform also runs an MCP server of its own, at https://mcp.mulno.com/mcp, so Claude Code or claude.ai can reach your sessions without installing anything. It answers from what Mulno's servers are allowed to read: a Mulno Cloud project. A Private project joins once its lent key is on, which is Soon. claude.ai has answered through it on production. ChatGPT has never been tried against it.

• To connect, add that address as a custom connector in claude.ai's connector settings, or in Claude Code run claude mcp add --transport http mulno https://mcp.mulno.com/mcp and sign in from /mcp. Either way you land on Mulno's own sign-in and consent pages, below. You need an account on the web app, and a member's grant only reaches the workspace and projects they tick.
• Sign-in is OAuth. The client registers itself (a client metadata document first, dynamic registration for older clients), and you sign in with a mailed link and give consent. The consent page shows the client's id and the host it will send you back to, since a client's own name proves nothing. PKCE is required and tokens are bound to the server they were issued for.
• A grant is one workspace and the projects you tick. Reading sessions is one scope; writing notes is a second, granted only if the client asked for it and you agreed.
• You can see and end your AI connections. Open /grants on the MCP server's own address, the same host that asked for your consent, and sign in with a mailed link. It lists each client you allowed, for which workspace and with what scopes, and Revoke ends one at once.
• Every call checks that you are still a member. A member who is removed is refused on the very next call.
• Nine tools, fewer than the local server: MCP tools. Anything a tester typed comes back marked as a tester's words, and each answer is cut at about 40 KB.