Script-Connected HTML Output

When an HTML file displays data from integrations, use data scripts so the data refreshes every time the file is opened. Only hardcode data when the source is truly static (user-provided CSV, pasted text).

Refer to the gumloop-sdk skill for the client and tool-calling patterns.

How It Works

For new script-connected HTML artifacts, start with:

python3 /home/user/skills/.tools/html_scaffold.py /home/user/myproject --script-connected

Then replace the generated data.py sample payload with the real Gumloop SDK call.

1. Write Python data scripts that fetch data via the Gumloop SDK and print JSON to stdout
2. Write the HTML to fetch('/gumloop/data/{key}') instead of hardcoding data
3. Export with sandbox_download(..., scripts={"key": "/home/user/script.py"}, server_ids=["server_id"])

The platform intercepts /gumloop/ fetches and runs the matching script server-side.

Example

Data script (/home/user/get_issues.py):

import json
from gumloop import Gumloop

client = Gumloop()  # base URL + token from env; project resolved from the token
resp = client.mcp.execute("linear", "list_issues", {"max_limit": 50})
print(json.dumps(resp.results[0].decoded_content))

HTML uses fetch to load that data and render it:

<script>
fetch('/gumloop/data/issues')
  .then(function(r) { return r.json(); })
  .then(function(data) {
    var container = document.getElementById('content');
    data.forEach(function(issue) {
      var div = document.createElement('div');
      div.textContent = issue.title;
      container.appendChild(div);
    });
  });
</script>

Export bundles both together:

sandbox_download(
  sandbox_path="/home/user/dashboard.html",
  scripts={"issues": "/home/user/get_issues.py"},
  server_ids=["linear"]
)

Keys in scripts and the fetch URLs must match. The tools each script may call are parsed from its source to scope the execute token — you don't list them. server_ids lists every server_id used by any script (enables connection checks when the artifact is shared).

Write Actions

For forms that write back to integrations, use /gumloop/action/{key} with POST. The request body is forwarded to the script as os.getenv("GUMLOOP_PAYLOAD").

import os, json
from gumloop import Gumloop

client = Gumloop()  # base URL + token from env; project resolved from the token
payload = json.loads(os.getenv("GUMLOOP_PAYLOAD", "{}"))
resp = client.mcp.execute("linear", "create_issue", payload)
print(json.dumps(resp.results[0].decoded_content))

Script Constraints

• Output JSON to stdout via print(json.dumps(...)) — stdout is the response body
• Instantiate Gumloop() directly (runs in a one-off sandbox)
• Scripts can only call tools from gumcp_server type servers (see server-discovery catalog). Do not use gumstack_server or mcp_server tools in data scripts — they will fail at execute time
• For read-only scripts, run them once in the sandbox before exporting to verify they return valid JSON
• Omit scripts from sandbox_download entirely for static HTML