How to run a creator program from ChatGPT, Claude or any MCP agent

You can run a creator program from an AI agent by connecting CollabPal as an MCP server. Add the server URL to Claude, ChatGPT, Codex, Cursor or any client that speaks Model Context Protocol, sign in once in the browser tab it opens, and the agent gets 85 tools that do exactly what the dashboard does: invite creators, organise them into groups, post bounties, review what comes back and pay approved rewards through Stripe. It acts with your permissions, scoped to the workspace you pick, and you can revoke it at any time.

This is the setup end to end, then the jobs that work well delegated, and the two you should keep a hand on.

How do I connect a creator program to an AI agent?

Add the server with no key and let the agent sign in. In Claude Code it is one command, then an authentication step:

claude mcp add --transport http collabpal https://www.collabpal.com/api/mcp
# then: /mcp  →  collabpal  →  Authenticate

The client opens a browser tab. You sign in to CollabPal, pick the workspace the agent may act in (or all of them), optionally make the connection read-only, and press Connect. The agent receives a token that refreshes itself. Any client that supports OAuth for MCP works the same way, which today includes Claude, ChatGPT, Codex and Cursor. For clients without a browser sign-in, or for scripts, create an API key under Account, API keys and pass it as a bearer header instead.

For any other client, the server block is:

{
  "mcpServers": {
    "collabpal": { "type": "http", "url": "https://www.collabpal.com/api/mcp" }
  }
}

Use the www host. The apex domain redirects, and most clients drop authentication on a redirect, which shows up as a confusing 401.

What can the agent actually do?

Everything the dashboard can, because the tools call the same server actions. The useful ones, in the order a program runs:

• Orientation: whoami, get_workspace. Ask the agent to start here so it knows which workspace and programs it is working in.
• Creators: invite_creators, create_join_link, list_creators, get_creator, add_creator_to_group, set_creator_status.
• Groups: create_group, list_groups. Groups are the audience a bounty is published to.
• Campaigns: create_campaign, publish_campaign, preview_campaign_reach, extend_deadline, close_campaign.
• Submissions: list_submissions, get_submission, review_submission, bulk_review_submissions, refresh_post_metrics.
• Payouts: list_payouts, pay_rewards, record_reward. pay_rewards charges the workspace's saved payment method and pays creators through Stripe.

There is also a creator side: an agent connected by a creator can list their programs, read a brief, submit work, check their earnings and connect socials. The same server serves both, keyed on who signed in.

Which jobs are worth delegating?

The ones that are frequent, well-defined and reversible. In practice that is most of the program's week.

• Onboarding: "Invite these twelve creators to the Launch group and send them the brief." The agent creates the group if it is missing, sends invites, and reports who accepted a day later.
• Posting: "Publish a bounty for a 60-second TikTok about the new feature, $150 per approved video, one per creator, open for two weeks, to the Launch group." The builder is one sentence of conditional logic, so the agent fills it in the same way you would.
• Reviewing: "Show me what came in overnight, approve anything that follows the brief, and request changes on the rest with a note." Ask it to list first and approve only after you have looked; approvals create money owed.
• Reporting: "Which creators have earned the most this month, and what did their posts do?" Every submission that links a post carries its views and likes, re-read on request, so the agent can answer with numbers rather than impressions.
• Housekeeping: extending a deadline, closing a bounty that filled up, resending an invite, moving creators between groups.

What should the agent not do on its own?

Two things move money or reach: paying, and publishing to a large group. Keep both behind a confirmation. The server's own instructions tell the agent to confirm before paying, deleting, removing creators or changing plans, and the downloadable skill file repeats it, but the reliable safeguard is you saying "show me the batch first".

• Before pay_rewards, ask for the list: who, how much, which submissions. The dashboard itself makes you hold the Pay button; give the agent the same pause.
• Before publish_campaign, ask for preview_campaign_reach: how many creators will see it and be emailed.
• Give the agent a read-only connection for reporting jobs. Read-only keys and sign-ins can look at everything and change nothing, including payouts.
• Scope the connection to one workspace unless you really want it working across all of them.

Do it with an agent

After connecting, paste this as the first message and the agent has everything it needs:

Use the collabpal MCP server. Start with whoami. I want to run a creator program for our launch next month: create a group called Launch, invite the creators in the attached list, and draft (don't publish) a bounty paying $150 per approved TikTok, one per creator, open for two weeks. Show me the reach before you publish anything, and never call pay_rewards without showing me the batch first.

Download the agent skill from Account, API keys as well. It teaches the agent the workflows and the things to confirm, so you are not repeating them each session.

Frequently asked questions

Does the agent get more access than I have?

No. Every tool runs under your own session and the same row-level permissions the dashboard uses. A member who cannot approve payouts in the dashboard cannot approve them through an agent either.

Can I see what the agent did?

Yes. Every review, invite, publish and payout is recorded the same way as a dashboard action, with your account as the actor, and shows in the submission history and payouts ledger. Connected apps are listed under Account, API keys, with last-used times.

How do I disconnect an agent?

Under Account, API keys, press Disconnect on the connected app. Its next call fails and it has to sign in again. API keys are revoked from the same page.

Which clients support the browser sign-in?

Claude (app and Code), ChatGPT, Codex and Cursor sign in over OAuth. Clients without a browser flow can use an API key as a bearer token; it is the same server and the same tools.

Can a creator use an agent too?

Yes. A creator who signs in gets the creator-side tools: their programs, open bounties, submitting work, earnings and socials. The server serves whichever side the signed-in account belongs to.