Connect an AI agent

AlphaProve's Model Context Protocol (MCP) connection is designed to let an AI agent use your strategies, backtests, and paper-trading simulation. Your agent needs support for Streamable HTTP and OAuth. It does not place live orders or move assets.

Availability: AlphaProve is in invite-only beta. Open the connection guide to check availability and choose your AI application, or copy the exact endpoint from Settings → Agent access. You need an eligible AlphaProve account to approve a connection.

Connect

1. Open Settings → Agent access in AlphaProve and copy the endpoint when it is available.
2. Add that address as a remote MCP server in your agent's connection settings.
3. Follow the sign-in link, review the application name, callback hostname, and requested permissions, then approve the connection.

If Agent access is unavailable, the endpoint has not been released or the server's operator has not enabled connections.

You grant access once for that connection. The agent refreshes its access token as needed; AlphaProve does not ask you to approve each tool call. Your agent may have its own confirmation settings, especially for paper-trading actions. Additional permissions require new consent.

Client setup

The /mcp page provides a current endpoint when discovery is available and tailors these same instructions to your client.

Claude

In Claude, open Customize → Connectors, choose Add custom connector, paste the AlphaProve endpoint, and select Add. Choose Connect to sign in and approve access, then enable AlphaProve from the Connectors menu in a conversation. See Anthropic's custom connector guide.

Claude Code

Run this in your terminal:

claude mcp add --transport http alphaprove <endpoint>

Start Claude Code, enter /mcp, and choose AlphaProve's Authenticate action. On versions with shell login, claude mcp login alphaprove starts the same OAuth flow. See the Claude Code MCP guide.

Codex

In the Codex CLI, run these two commands with the endpoint copied from Settings → Agent access:

codex mcp add alphaprove --url <endpoint>
codex mcp login alphaprove

The second command opens the OAuth sign-in flow. Review the requested permissions before approving. See the current Codex MCP documentation.

ChatGPT

ChatGPT uses a manual connection setup. First turn on Developer mode in Settings → Security and login, then open ChatGPT Plugins and select the plus button. Enter AlphaProve as the name, add a short description, and under Connection enter the public MCP endpoint, then create the connection and review the discovered tools. In a new conversation, add AlphaProve from the tools menu and complete the AlphaProve sign-in and consent flow.

Developer mode and full MCP availability depend on your account, plan, and workspace policy. Ask your workspace administrator if these controls are unavailable. See OpenAI's current Connect and test your plugin guide.

Other clients

Add the endpoint as a remote Streamable HTTP server in the client's MCP settings. Choose OAuth, complete the AlphaProve sign-in, and review the requested permissions. Client labels and availability vary, so follow the client's current MCP documentation if its controls use different names.

Choose permissions

Permission	What the agent can do
Read account research	Read your strategies, runs, results, SDK references, allowances, and available market summaries. Validate strategies and preview research.
Create and test strategies	Create, update, and duplicate strategies; start and cancel backtests, forward walks, and Lab campaigns; calculate risk evaluations.
Manage paper trading	Start, pause, resume, and stop paper runs.

Read access is required alongside either of the other permissions. Paper access can manage existing strategies without permission to edit them.

Your agent shares your account's limits and credits. Website and agent work use the same allowances, subscription features, and concurrency limits. There is no separate budget per connection. A preview is an estimate; availability and cost are checked again when work is submitted.

Write and test a strategy

Try asking your agent:

Read AlphaProve's authoring rules, complete SDK contract, and examples. Check what market data and account features I can use. Create a simple strategy, validate it, and backtest it on a suitable date range. Explain the assumptions, trades, drawdown, and result, and give me the AlphaProve link.

A reliable agent workflow is:

1. Call get_capabilities and get_market_catalog.
2. Read get_reference topics authoring, sdk, examples, and run_settings. Read the returned named sections when a topic spans multiple sections.
3. Call validate_strategy, correct any errors, then create_strategy.
4. Call start_backtest. Keep the returned run ID and website link.
5. Poll get_run at the suggested interval. Once it finishes, read get_run_results, get_run_trades, and get_run_equity.

The reference includes the versioned Python contract, history requirements, sandbox rules, examples, and settings. Agents should use it before writing code. The same material is available through alphaprove://reference/authoring, alphaprove://reference/sdk, alphaprove://reference/examples, and alphaprove://reference/run_settings for clients that support MCP resources.

To edit an existing strategy, first read it with get_strategy. Pass its current revision to update_strategy. If someone edited it in the meantime, the update returns a revision conflict: read the latest strategy and reconcile the edits. Existing runs retain the strategy snapshot they used.

Other workflows

Workflow	Tools
Find and review strategies	list_strategies, get_strategy, list_builtin_strategies, duplicate_strategy
Compare backtests	compare_backtests
Walk-forward research	inspect_forward_walk, start_forward_walk
Lab research	preview_lab, start_lab
Risk and prop-firm evaluation	evaluate_monte_carlo, get_prop_firm_catalog, evaluate_prop_firm
Paper trading	start_paper_run, control_paper_run
Track work and results	list_runs, get_run, get_run_results, get_run_trades, get_run_equity
Cancel research	cancel_run
Market context	get_market_summary

The tool schemas describe supported run types, result sections, filters, and page sizes. get_capabilities explains which tools your connection and account can currently use. Follow pagination information to retrieve more results.

Retries and running work

For each mutation, generate a new request_id. If a response is lost, retry the same call with the same request ID and unchanged arguments. AlphaProve keeps a receipt so a retry can recover the accepted result. Reusing an ID for a different action returns an idempotency conflict. get_operation checks the request's status for the same account and agent application.

Backtests, forward walks, Lab campaigns, and paper runs continue if the agent disconnects. Use cancel_run to cancel research. Use control_paper_run to pause, resume, or stop paper trading. Cancelling a completed research run preserves its saved results.

Disconnect or reconnect

Open Settings → Agent access and select Revoke on the connection. It blocks further calls and token refreshes. Already accepted work continues; use the website to manage it. Signing out of the website normally leaves agent connections active. Log out everywhere invalidates their credentials too.

If your agent reports expired or revoked access, reconnect through OAuth. For a missing-permission error, reconnect with the permission needed for that action. For a quota or plan restriction, inspect get_capabilities and your account's current allowance before submitting again.

Acceptance status

Local protocol smoke tests have exercised the Python MCP SDK and mcp-remote against a local endpoint. Those checks verify the protocol boundary only; they do not establish acceptance in Claude, Codex, ChatGPT, or another hosted client. Real-client validation remains pending and must cover the current client version, public HTTPS reachability, OAuth approval and denial, tool discovery, historical backtesting, paper simulation, reconnect, token refresh, and revocation before client support is claimed.