Run your affiliate program from Claude or ChatGPT.

Why run a partner program from an agent at all?

A dashboard is good for looking at an affiliate program, and slow for working on one. A typical week is two minutes of judgement buried under half an hour of clicking: open each application, check the applicant's site, approve, move the good ones into the right reward group, answer a partner asking when they get paid, then export commissions for the accountant at month end.

Those are exactly the jobs a language model with tool access does well. The judgement stays yours. The agent collects the context, proposes the decision and carries it out once you agree.

Connecting the agent

Partli hosts its MCP server at one URL. Most clients sign in with OAuth, so you don't need to copy a key:

# Claude Code
claude mcp add --transport http partli https://partli.app/api/mcp

# Claude.ai or ChatGPT: Settings → Connectors → add custom connector
#   URL: https://partli.app/api/mcp

# Cursor (.cursor/mcp.json)
{ "mcpServers": { "partli": { "url": "https://partli.app/api/mcp" } } }

If you run several products, each one is its own workspace. Point a connection at one with/api/mcp/<workspace-slug>, or pass workspace on any tool call. Headless setups (CI, scheduled agents) use a workspace API key instead of OAuth. The MCP and CLI docs have every option.

What the agent actually sees

Here is a real call from our own workspace. Partli runs its affiliate program on Partli. Asking “is our partner program set up properly?” makes the agent call setup_status, which returns the go-live checklist (trimmed):

{
  "workspace": "partli",
  "checklist": [
    { "step": "Create your first program",        "done": true },
    { "step": "Install tracking on your site",    "done": true },
    { "step": "Verify tracking end-to-end",       "done": false,
      "fix": "verify_attribution — the user opens the test URL and completes a checkout" },
    { "step": "Sales connected (Stripe, Paddle/Polar/Creem/Dodo, Shopify/WooCommerce, or the API)", "done": true },
    { "step": "A program is live",                "done": true },
    { "step": "Approve your first partner",       "done": true },
    { "step": "Connect your AI agent (MCP)",      "done": true }
  ],
  "partnersByStatus": { "approved": 5, "rejected": 1 }
}

Every unfinished step comes with a fix naming the next tool to call. The agent doesn't have to guess what to do next, and it can tell you plainly what's missing: here, nobody has yet run an end-to-end test purchase.

Jobs worth handing over

Where to keep a human in the loop

• Money. run_payouts, mark_manual_payouts_paid and similar tools carry MCP's destructive annotation, so clients that respect it ask before calling. Keep that setting on.
• Bans and rejections. Let the agent propose them with its reasons, then approve them in a batch.
• Messages to partners. Have the agent draft and show you broadcast emails before send_campaign.
• Scope. Give automated, unattended agents a workspace API key for one product, not your whole account.

Widgets in Claude and ChatGPT

Claude and ChatGPT support MCP Apps, which let a tool return a small interactive interface alongside its text. Partli ships two: an application-review list with approve and reject buttons, and a payouts summary grouped by currency. In text-only clients such as Codex or a terminal, the same tools return structured text, so nothing is lost.

Which affiliate platforms support AI agents?

MCP support spread quickly through affiliate software in 2026, but the servers differ a lot in what they let an agent do. Here is what each vendor publishes as of October 2026:

From each vendor's own website, checked 17 October 2026.

If you only want to ask questions of your data, any of the read-only servers will do. If you want the agent to carry out the work (approving people, changing rewards, closing the month), check that its tools can write. Our guide to affiliate software for SaaS and the head-to-head comparisons cover the rest of each product.

FAQ