# Kranth for agents

Kranth runs a room of synthetic customers with real biases against the thing your user is about to spend money on, before they spend it. Every version is judged by the same pinned room. The result is a decision (keep, fix or kill each version), the objections behind it, who raised them, and a fix list.

Connect: add `https://mcp.kranth.ai/mcp` as a connector. The user signs in at kranth.ai, picks the workspace (the brand whose audience and company brief shape every run), a monthly credit cap, and what you may do.

## Pick the tool

| The user wants to test | Call |
|---|---|
| 2 to 5 headlines, hooks, taglines, subject lines as plain lines | `kranth_ab_test` |
| 1 to 5 ad images | `kranth_check_creatives` (upload local files first with `kranth_upload_image`) |
| a full ad unit: image + headline + body + CTA | `kranth_run_simulation` type `full_ad` |
| a landing page | `kranth_run_simulation` type `page`, one variant `{url}` |
| store listings (Amazon, Shopify, App Store) | `kranth_run_simulation` type `listing` |
| pricing tiers, offers, guarantees, promos | `kranth_run_simulation` type `offer` |
| cold emails | `kranth_run_simulation` type `cold_email` |
| an idea, feature, plan, policy | `kranth_run_simulation` type `idea` |
| a decision argued out between personas | `kranth_run_debate` |
| the user being questioned on a decision | `kranth_start_interview`, then `kranth_interview_reply` per answer |

## The loop

1. `kranth_account` once: workspace, permissions, credits left.
2. Optional `kranth_quote` for a cost before a big run.
3. Start the run. You get a `run_id` back in seconds.
4. `kranth_get_run` with `wait_seconds=45` until `status` is `complete` (usually 1 to 4 minutes).
5. Tell the user, in plain words: the decision, the winner, each version's verdict and top objection, the best quote, and the fix list. Link `open_url` for the full report.
6. Offer next steps: `kranth_improve` for a rewrite to re-test, `kranth_set_public` for a share link, `kranth_export` for the PDF.

## Rules

- **Pre-flight, not prediction.** Kranth surfaces objections before spend. Never say it predicts CTR, conversions or revenue.
- **Image runs need a model that sees images.** `kranth_list_models` marks them `sees_images: true`. The default model sees images.
- **Interviews are the user's words.** Relay each panel question verbatim and send the user's real answer. Never answer for them.
- **Ask before publishing.** `kranth_set_public` makes a result readable by anyone with the link.
- **Errors are sentences.** Credit, plan, cap and permission errors come back as plain text. Relay them as they are; they say where to fix it.

## Prompts to try

- "Test these three headlines on B2B SaaS buyers and tell me which one to ship: …"
- "Here are two ad images for our water bottle. Which one should I put money behind, and what's wrong with the other?"
- "Run our pricing page through Kranth and tell me where people drop off."
- "Which of these two cold emails gets more replies from VPs of Marketing?"
- "Interview me about raising prices 40% next month. I want to know if I'm ready."