Ship it to a human. Get structured feedback back.

Terrarium is the review layer you can call. Publish HTML, and a human reviews it in the browser — comments pinned to the exact words, approve or request changes. You pull that verdict over MCP as a typed payload, fix, and push a new version. Loop until approved.

$ npx skills add nelsonsantryhkd/terrarium-plugin

Installs the MCP server + the skill that teaches the loop. Then mint a token at tryterrarium.com/connect as TERRARIUM_PAT.

Four moves. Two roles. One closed loop.

Every step is a tool you already have. The human half happens in the browser — you never see it, you just get the structured result.

1. Publish (agent) — one publish_artifact call turns raw HTML into a hosted, sandboxed page with a share link.
2. Review (human) — highlight the exact words, leave a typed comment — change request, question, or bug — then approve or request changes on that version.
3. Pull & fix (agent) — pull_comments returns the verdict and every thread anchored to its passage — structured, not prose.
4. Push (agent) — push_version ships the fix; comments carry forward, re-anchored. Back into the review queue.

Feedback as a contract, not a chat log

pull_comments returns structured output alongside markdown: the review verdict, typed threads in document order, and text anchors with live status — anchored, needs-review, or orphaned.

The payload is versioned and evolves additively within terrarium-feedback/1. Breaking it is a new protocol version, announced — because agents get trained against this format.

{
  "protocol": "terrarium-feedback/1",
  "review": {
    "state": "changes_requested",
    "note": "Start the axis at zero."
  },
  "expected_version": "41c0…",           // pass to push_version
  "threads": [{
    "kind": "change_request",
    "status": "open",
    "body": "This reads as 5× growth.",
    "anchor": {
      "type": "text",
      "quote": "axis starts at 12%",
      "status": "anchored"
    }
  }]
}

Your agent sets itself up

1. Install the skill + MCP server — one command bundles the tools and the loop skill.

   $ npx skills add nelsonsantryhkd/terrarium-plugin
   /plugin marketplace add nelsonsantryhkd/terrarium-plugin

2. Mint a token — the one step that needs a human. Grab a PAT at tryterrarium.com/connect and export it as TERRARIUM_PAT. You act as that user, under their permissions.
3. Run the loop. Publish, hand over the link, then pull_comments until the verdict is approved.

Built for machines to call

• Seven MCP tools over Streamable HTTP: list, publish, get, pull comments, push version, reply, resolve.
• Stale-push protection — every payload carries expected_version; racing pushes fail loud, never clobber.
• A versioned contract — terrarium-feedback/1, additive-only, safe to train against.
• Every call logged to your token — the human can see exactly what their agent did.

nelsonsantryhkd/terrarium-plugin — the Terrarium plugin for Claude Code: MCP server config, the loop skill, and install docs. Open source · MIT.

Prefer no code? See the human, drag-and-drop flow →

Agents build. Humans steer.

Give your agent a place to show its work — and a verdict it can act on. Install the skill, mint a token, run the loop.