# A/B testing for AI agents: TryFast MCP server

Let your AI agent run your A/B tests. You keep the launch button.

Connect TryFast to Claude, ChatGPT, Cursor or VS Code with one URL. Your agent installs the snippet, looks at your pages, builds experiments, sends you preview links and reads the results. Nothing goes live without your approval unless you allow it.

```bash
claude mcp add --transport http tryfast https://tryfast.eu/mcp
```

[Connect your agent](https://tryfast.eu/app/zero-account-setup?next=agents) or [read the MCP docs](https://tryfast.eu/docs/mcp.md). No account needed to start. You approve launches. Hosted in Germany.

Connecting creates a free workspace and accepts our [Terms of Service](https://tryfast.eu/legal/terms-of-service.md) and [Data Processing Agreement](https://tryfast.eu/legal/data-processing-agreement.md).

## Connect in one step

The MCP server URL for every client is `https://tryfast.eu/mcp`.

- **Claude Code**: `claude mcp add --transport http tryfast https://tryfast.eu/mcp`, then type `/mcp`, choose tryfast and Authenticate. For a team: add `--scope project` to write `.mcp.json`.
- **Claude** (web, desktop, mobile): Customize > Connectors > Add custom connector. Name: TryFast. URL: https://tryfast.eu/mcp. Then Connect and approve in the window that opens.
- **ChatGPT**: Settings > Apps > Advanced settings: turn on Developer mode. Create app, name TryFast, MCP server URL https://tryfast.eu/mcp, authentication OAuth.
- **Cursor**: the "Add to Cursor" button on https://tryfast.eu/ai-agents, or add `{"mcpServers":{"tryfast":{"url":"https://tryfast.eu/mcp"}}}` to `~/.cursor/mcp.json`.
- **VS Code**: `code --add-mcp '{"name":"tryfast","type":"http","url":"https://tryfast.eu/mcp"}'`, or open https://tryfast.eu/go/vscode-mcp.
- **Other agents and CI**: any MCP client with OAuth works. Without a browser, create an API key and send it as `Authorization: Bearer`.

When you connect, a TryFast window asks which workspace the agent may use and how much it may do. No account yet? Choose "Continue without an account".

## What your agent can do

- **Set up**: create a project, get the right snippet for your stack, and confirm it is installed (`create_project`, `get_install_snippet`, `check_installation`).
- **Look**: read the structure of a page to find stable selectors, and see live events to check that goals fire (`inspect_page`, `get_events`).
- **Build**: create experiments with changes, CSS, JavaScript or redirects, define goals, and share preview links (`create_experiment`, `update_experiment`, `create_goal`, `get_preview_link`).
- **Run**: launch, pause, change traffic and stop, with your approval where you want it (`start_experiment`, `pause_experiment`, `set_traffic_allocation`, `stop_experiment`).
- **Learn**: read results with a verdict, estimate how long a test needs, and search the docs (`get_results`, `calculate_sample_size`, `search_docs`).

## An example session

> **You**: Test a free shipping headline on acme-shop.de/pricing against the current one. Goal: more people starting checkout.
>
> **Claude** (`inspect_page`, `create_experiment`, `get_preview_link`): The headline is `#hero h1`, "Simple plans for every shop". I created the draft "Free shipping headline": Variation B reads "Free shipping on every order". Goal: Started checkout. Preview B on your site: acme-shop.de/pricing?tf_preview=... At about 4,500 visitors a day, detecting a 10% lift takes about four weeks; a larger effect shows sooner.
>
> **Claude** (`start_experiment`, awaiting approval): Ready to launch. Approve it here: tryfast.eu/app/approve/apr_8f2kQ7xLm1
>
> **You** (two weeks later): How is the free shipping test doing?
>
> **Claude** (`get_results`): Variation B converts better: +15.0% more visitors start checkout (95% interval +1.0% to +31.3%, adjusted for two variations). It has been safe to act on since day 10. Variation C shows no clear difference. Shall I ship B? That needs your approval: tryfast.eu/app/approve/apr_q71mT3vB9c

## Built so an agent can help without surprises

- **Approval links.** With the default permission, launches, traffic changes and stops wait for one click from you. You see exactly what goes live.
- **Three permission levels.** Read only, Build with your approval, or Full access. Change or revoke them any time.
- **Pausing never waits.** An agent with Build or Full access can pause an experiment immediately, without approval.
- **Every action is logged.** The activity log shows the agent's name and the person it acts for next to every change.
- **Statistics safe to poll.** Agents check results often. Sequential statistics keep those checks from turning noise into winners, and the verdict says whether it is safe to act.
- **Page content stays data.** Text from your pages and events reaches the agent as labelled data, never as instructions.

## Start from inside your agent

Your agent can start before you have an account. In the connection window, choose "Continue without an account" and TryFast creates a free workspace for you. Save it to an account later and everything stays.

## For agents without a browser

Create a personal API key in the app and send it as a header. Keys have the same three permission levels.

```json
{"mcpServers":{"tryfast":{"type":"http","url":"https://tryfast.eu/mcp","headers":{"Authorization":"Bearer ${TRYFAST_API_KEY}"}}}}
```

Agents with only HTTP can create a workspace with one request to the API and hand you a link to save it. Launches still wait for you: open the link, sign in and approve. See the [API docs](https://tryfast.eu/docs/api.md).

## Starter prompts

- Install TryFast on acme-shop.de and check that it works.
- Look at https://acme-shop.de/pricing and propose three A/B tests, ranked by expected impact and effort.
- Create an A/B test for a shorter checkout button text, with "Started checkout" as the goal, and send me preview links.
- How are my running experiments doing? Tell me which ones are safe to act on.

## Questions about agents

### What is MCP?

The Model Context Protocol is an open standard that lets AI agents use tools. TryFast's MCP server gives your agent the same abilities as the TryFast app.

### Which agents work with TryFast?

Claude Code, Claude, ChatGPT (developer mode), Cursor, VS Code and any other client that supports remote MCP servers with OAuth.

### Can the agent see my visitors' personal data?

Agents see aggregated results and recent technical events with random, pseudonymous visitor ids, never IP addresses, the same as you see in the app. Do not put names or email addresses into event properties.

### What does an agent cost?

Nothing extra. Agent access is included in every plan, including Free.

### How do I disconnect an agent?

In the app under AI agents, choose Revoke. The agent loses access immediately.

## Give your agent a testing tool it can be trusted with

[Connect your agent](https://tryfast.eu/app/zero-account-setup?next=agents) or [read the MCP docs](https://tryfast.eu/docs/mcp.md).
