Your AI can now run AlphaProve
Create strategies, run backtests and inspect the results from your AI app. How the MCP connection works, and the small callback bug we caught before release.
An AI writes you a trading strategy. You copy it into a backtester, fix the parts that don't fit, run it, then paste the results back into the conversation. The next revision starts the same round trip.
AlphaProve now has an authenticated MCP connection that removes that handoff. Your AI agent can write a strategy, check it, submit a backtest and read the results in your account. You can carry on the conversation while the work runs on AlphaProve.
MCP stands for Model Context Protocol. It's the interface AI applications use to call external tools and read their documentation. Our connection uses Streamable HTTP and OAuth. You add the server to your AI app, sign in to AlphaProve and choose its permissions.
There are setup guides for Claude, Claude Code, Codex and ChatGPT, plus a generic guide for other compatible clients. Client support still depends on your AI application's plan and settings. AlphaProve is in invite-only beta, so you'll need an approved account.
A research session without the copy-paste
Here's an example request you could give a connected agent:
Read AlphaProve's strategy SDK and check which data and account limits are available. Help me test a BTC trend-following idea on one-hour candles. Ask me about the entry, exit and position size before writing it. Validate the strategy, run a backtest, then show me the trades and drawdown with a link to the full result.
Before choosing a test window, the agent can check your account's allowances and available market data. It saves the strategy in your library, validates the code and submits the run. Once the run finishes, it can read the summary, individual trades and equity curve.
You can then ask it to compare two saved backtests, inspect the trades around a drawdown, or explain which settings changed between runs. The result links open the same records in AlphaProve, so you can inspect the chart yourself.
Your strategies and runs stay in your AlphaProve account, available on the website after you leave the conversation.
What the connection exposes
The first release has 29 tools and four reference resources. The full guide covers their setup and limits; these are the jobs they support.
| What you want to do | What the agent can use |
|---|---|
| Prepare a test | Account permissions and allowances, available markets, market summaries, builtin strategies and SDK references |
| Work on a strategy | Create, read, update, duplicate and validate strategies |
| Run a backtest | Submit a run, check its status, cancel it and read its results, trades and equity curve |
| Compare and investigate | Compare backtests, run Monte Carlo analysis and evaluate results against supported prop-firm rules |
| Continue the research | Start walk-forward tests and inspect their batches; preview, start and cancel Optimizer Lab campaigns |
| Try paper trading | Start a session, pause or resume it, stop it and read its results |
Your account's permissions and plan limits apply to every tool, so the features you can use depend on your plan.
Paper trading uses simulated money. The MCP connection cannot place live orders or move funds. Disconnecting an app also doesn't stop a paper session it already started; stop that session separately if you want it to end.
The agent needs the SDK, too
An AI using create_strategy needs to know what valid AlphaProve code looks like.
Plausible Python can call a method that doesn't exist, request history incorrectly or assume the wrong timing between a signal and an order. Successfully uploading a strategy doesn't establish that those details are right.
We exposed the complete strategy SDK reference alongside the authoring rules, examples and run settings. Agents can read the reference by topic and section. The connection page also provides a starter prompt that asks the agent to read every returned section before it starts writing.
For example, AlphaProve strategies declare the history they need through HISTORY. The agent should read that requirement before drafting code. Making it discover the rule through failed submissions would waste everyone's time.
Validation checks the strategy before execution, and user strategy code runs in AlphaProve's server sandbox. The AI doesn't need database credentials or a copy of the market-data infrastructure to run a test.
Why we kept the existing backtesting machinery
The MCP tools call our existing services for saving strategies, applying account limits and submitting work to the workers. AI requests use the same simulation path as the website.
A backtest submitted through an agent therefore runs on the same machinery as one submitted through the website. Its results belong to the same account, and its usage comes out of the same allowances.
Long-running work returns a run handle that the agent can check later. If a client loses a response and retries with the same request ID, AlphaProve returns the recorded operation. That prevents the retry from submitting the job twice.
The release bug was spelled localhost
Our protocol tests passed, and production checks exercised every exposed tool. Backtests and a walk-forward run completed; we also checked the paper-trading and Lab lifecycles. But when we tried the actual Claude Code client, it failed before the sign-in page opened.
Claude Code registered a callback like http://localhost:52772/callback. Our OAuth validation allowed local HTTP callbacks for 127.0.0.1 and ::1, but rejected the hostname localhost.
We changed the validator to accept that exact hostname as a local callback. It still rejects remote HTTP addresses and applies the same redirect matching and PKCE checks. PKCE ties the authorization-code exchange to the client that started the sign-in.
We added a regression test for the callback Claude Code used, along with checks that lookalike hostnames and ordinary remote HTTP URLs remain rejected. After deployment, Claude Code completed browser approval and reported the connection as connected.
We couldn't complete an AI-driven conversation in that test: our Claude organization had disabled subscription access for Claude Code, even though MCP sign-in succeeded. The official MCP client could read the SDK using the approved grant. The full Claude Code conversation remains on our client test checklist.
Our own client had used an IP address throughout testing. It took the application people would use to expose the hostname bug.
You choose the access
AlphaProve asks for permission once per connection. The permissions separate reading account research, creating and testing strategies, and managing paper trading. Adding permissions requires another approval. Your AI app may also ask you to confirm individual actions.
You can review or revoke a connection in Settings → Agent access. Choose an AI app you're comfortable sharing your research with: it can receive the account data you've allowed, including strategy source and results.
The same AlphaProve credits and run limits apply across the website and MCP. Your AI provider's own subscription or API charges are separate.
Connect and try one idea
Open Connect your AI, choose your application and follow its setup guide. For Claude Code, the server command is:
claude mcp add --transport http alphaprove https://alphaprove.com/api/v1/mcp
Then enter /mcp inside Claude Code and authenticate AlphaProve in your browser. The setup documentation includes the other clients and the current authentication options.
Start with a strategy you can explain clearly, and ask the agent to read the SDK before writing it. Have it validate the code and return a result link you can inspect. The connection saves the work of moving code and results between apps. You'll still need to judge the assumptions in that historical simulation before risking money.