MCP server

Versuch exposes a Model Context Protocol (MCP) server so an AI assistant can set up and manage experiments, pricing tables, and feature flags, and read usage, directly from your editor or chat. No dashboard needed.

Endpoint

https://versuch.ai/api/mcp

• Auth: Authorization: Bearer vsk_live_… (your project secret management key)
• Protocol: MCP 2024-11-05 over JSON-RPC (HTTP POST)
• Server name: versuch

Caution

The vsk_live_… key is a secret management key. Keep it in your MCP client config / a secret store, never in client-side code.

Client config

Add Versuch as a remote MCP server. Replace YOUR_SECRET_KEY with your vsk_live_… key.

mcp config

{
  "mcpServers": {
    "versuch": {
      "url": "https://versuch.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_SECRET_KEY"
      }
    }
  }
}

Tools

The server exposes 12 tools. Reads (list_*, get_*) are always available; the mutating tools are plan-gated (the free tier is watch-only).

Tool	What it does
list_experiments	List all A/B experiments (status, goal, target selector, variants).
create_experiment	Create and (by default) activate an experiment. type element swaps a selector’s text, cta swaps a button’s label/link/colors, manual has no DOM effect (assign + emit goals yourself). 2–6 variants, one keyed control.
set_experiment_status	Start, pause, or stop an experiment by id (running / stopped / draft).
get_experiment_results	Per-variant exposed / conversions / rate, uplift %, p-value, significance, and the winner.
list_pricing_tables	List pricing tables (payment cards) with status, slot, and card count.
create_pricing_table	Create and (by default) activate a pricing table rendered into <div data-versuch-pricing="slot">. Standalone cards or A/B variants; a card may carry a Stripe price_id.
get_usage	Account event usage this billing period vs the plan limit, with a per-site breakdown.
get_install_snippet	The tracking <script> snippet and public ingest key for the project.
list_flags	List feature flags (key, name, type, tags, temporary, archived).
create_flag	Create a flag. boolean (default) is on/off; string / number / json are multivariate (supply variations). Created disabled in every environment.
set_flag	Enable or disable a flag in one environment (default production) and republish to the edge. Use enabled: false as a kill switch.

Tip

create_experiment, set_experiment_status, create_pricing_table, create_flag, and set_flag mutate and are plan-gated. If a call is rejected, the tool error explains why and what to do.

Example: using it with Claude

1. Add the config above to your MCP client (Claude Desktop, Claude Code, Cursor, …) with your vsk_live_… key.

2. Ask in natural language, e.g.:

   Create an A/B test on my homepage hero headline. Test “Ship 2x faster” against the current copy, and use the signup event as the goal.

   The assistant calls create_experiment with type: "element", your selector, and two variants, then activates it.

3. Follow up to read results:

   How is the hero headline test doing?

   The assistant calls get_experiment_results and reports per-variant rate, uplift, and significance.

Related

For installing the tracker and general integration guidance aimed at AI coding agents, see llms.txt.