ZUMA / AGENT LABSynthetic data only

Can Grok run our jobs?

A small test of the route from a CRM task to a browser action and a checked result.

Test queue
Approved task
→
Grok Bot
Browser action
→
Independent check
Actual message

Ready to connect. Local checks and live Grok runs are tracked separately. The operator results contain the current test status.

Connect your Grok Bot
  1. Add a custom remote HTTPS MCP connector at https://agents.zumahq.com/mcp. Choose OAuth / browser sign-in. Do not supply a manual token.
  2. Open the sign-in link yourself. Use the invented grok-a login from the local .runtime/setup.html file, then click Sign in and allow test access.
  3. Keep login input outside agent observation. The connector receives its own scoped token directly.
  4. Before the first job, ask Grok to open this site's /login page and let you take over privately. Sign into the fake messaging app, then return control.
  5. Paste the task below into Grok. A second tenant requires a separate Grok user account and its own login.

The connector sign-in and messaging-app sign-in are separate connections. Both sign-ins worked in the founder's synthetic test.

The task to paste into Grok
Run one Zuma synthetic messaging test using the connected MCP tools.
Call claim_test_job once. If no job is returned, stop.
Open the returned browser_url using your computer browser. Do not use the terminal, HTTP clients, or hidden page state to send messages.
If login is needed, ask me to take over and sign in privately. Report needs_human with the job ID and lease, then stop. Never ask for a password in chat.
Select the exact recipient and enter the exact supplied text. Click Send once.
If the UI confirms the sent recipient and text, report sent. If sending is uncertain, report outcome_unknown. Do not retry Send.
Use report_test_result with job_id, lease and outcome. Do not claim another job this run.
Only interact with this synthetic test site. Do not open real LinkedIn or other accounts.
What counts as a pass?

Ten correct messages across two separate Grok accounts, with no duplicate sends. Expired login, lost confirmation, revoked access and cross-account attempts must produce the right result. Queue checks alone do not establish vendor isolation or password exclusion.

First run one task manually. Test a short, bounded routine only after that works. Record latency, handoffs and vendor usage separately.

Privacy boundary of this test

This server stores only invented accounts. Credentials are excluded from MCP tool responses. Testing Grok's private login flow and inspecting its available traces remain live checks. No claim is made about vendor-internal payloads.