wonder

wonder

A project from the AI for Science Power Users gathering.

wonder helps the people at the gathering find each other. You've been working in a Claude session on some piece of science. You run /wonder. It reads what you've been doing, writes a short, shareable summary of what you're wondering about — plus a few keywords — scrubbed of anything sensitive. Once you approve it, it hands those keywords to OhWow, a service that matches you against everyone else and drops you into a Discord room with the people working on the same thing.

Status: working MVP, deployed. /wonder runs end-to-end against a hosted OhWow MCP server (on Railway), creating/joining real Discord rooms. Install it from ohwow.science. See the live walkthrough just below.

The idea

The gathering is full of people quietly working on overlapping problems who will never discover each other in a hallway. But each of them has just spent hours explaining their problem to Claude. That session is the perfect description of what they're working on — it's just also full of things they'd never want to broadcast (unpublished results, file paths, API keys, collaborators' names).

wonder turns that session into a safe set of keywords (and a one-paragraph "here's what I'm wondering about"). OhWow takes those keywords, finds the other people wondering about the same things, and connects them in a Discord room — seeded with those shared keywords so there's an instant starting point for conversation.

The feeling we're after: you're heads-down in a session, a little lonely talking only to agents, you type /wonder — and a few minutes later you're in a room with humans who care about the same thing you do.

Naming

• /wonder — the command a user runs inside Claude Code.
• OhWow — the MCP service everyone's /wonder talks to. It matches on keywords and owns the Discord room creation + invites. (Working name; candidate domain TBD — ohwow.science was floated.)

Live walkthrough

A real /wonder run against the hosted OhWow MCP, using a Claude session as the source. The five steps the command performs:

1. Summarize the session → a wonder profile. Distill the live session into a one-line "wondering about", field, methods, and a tight keyword set. Example from a session spent building OhWow itself:

   • Wondering about: wiring Claude Code sessions to a Discord matchmaking service over MCP
   • Keywords: mcp server, discord api, claude code plugins, researcher matchmaking, ai for science

2. Scrub. Run the scrub skills over the draft. In this run, none of the secrets / file paths / tokens that appeared in the session made it into the keywords.

3. Review gate. Show the user the exact keywords. Nothing is sent until they approve.

4. Connect via the OhWow MCP. suggest_rooms previews matches (read-only), then connect does the one side-effecting call:

   join_demo            -> demo mode active? false
   suggest_rooms        -> [{room: serendipitous-matching-roundtable, overlap: 1},
                            {room: discord-invite-links-lab, overlap: 1}]
   connect              ->
     room:            #serendipitous-matching-roundtable
     reused existing: true (overlap 1)        # convergence: routed into a live room,
                                              # not a brand-new empty one
     serverInviteUrl: https://discord.gg/…    # join the server (if you're not in it)
     roomUrl:         https://discord.com/channels/…/…   # then open the room
   

5. Hand off. The user gets two links — serverInviteUrl to join the server and roomUrl to open the room. Already-members just use roomUrl.

The notable bit: connect reused an existing room instead of creating a new one, because the keywords overlapped — so overlapping topics converge rather than multiply. Rooms are only ever created/joined by an explicit, user-triggered /wonder; nothing runs in the background except a delete-only idle-room reaper.

How it works

flowchart TD
    S["Your Claude session"] -->|"/wonder"| SUM["Summarize: what am I wondering about? + keywords"]
    SUM --> SCRUB["Scrub sensitive info: secrets, PII, unpublished science, local env"]
    SCRUB --> REVIEW{"You review and approve the exact text + keywords"}
    REVIEW -->|"edit / redo"| SCRUB
    REVIEW -->|"approve"| OHWOW["OhWow service"]
    OHWOW --> MATCH["Match people on shared keywords"]
    MATCH --> ROOM["Programmatically create a Discord room, seeded with the keywords"]
    ROOM --> INVITE["Each person gets an invite link — you click to join"]
    ID[("ORCID sign-in (stretch / bonus)")] -.-> OHWOW
    VAL[("Valency: researcher metadata (stretch)")] -.-> MATCH