Communitygithub.com

algorover-research

Discover AlgoRover Research strategies, preview and run historical backtests, inspect or cancel jobs, and download research reports through the hosted MCP service.

What is algorover-research?

algorover-research is a Claude Code agent skill that discover AlgoRover Research strategies, preview and run historical backtests, inspect or cancel jobs, and download research reports through the hosted MCP service.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/algorover/algorover-research-skills/tree/main

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

AlgoRover Research

Use the host's algorover-research MCP connection at https://research.algorover.ai/mcp. Installing this Skill adds instructions; configure the MCP connection separately using connection instructions when its tools are missing.

Research uses public historical market data. It does not connect exchange accounts or execute orders. Public discovery and preparation do not require sign-in. Personal runs and reports require browser OAuth with research:read and research:run; never request credentials in chat.

  1. For discovery, call search_strategies with locale: "zh-CN" for Chinese and locale: "en" for English or other languages. Preserve that locale in later tools that accept it. For a named strategy, use the returned reference to call get_strategy.
  2. Use only the version, presets and markets returned by the service. If preparation is not_requested, call prepare_strategy; check get_strategy at least three seconds apart until ready. Report its failure reason when unavailable.
  3. Call prepare_backtest for the selected preset, venue and supported 90- or 365-day window. Present the exact configuration and wait for confirmation before starting.
  4. After confirmation, call start_backtest using the returned prepared_input, preparation_ticket and request_key as idempotency_key. Reuse that key if retrying an uncertain start. Preserve the real task identifier.
  5. Use get_backtest for status, get_backtest_result for completed metrics and get_backtest_report_download for requested reports. Use list_backtests when the task identifier is unknown. Show actual returned text and links; mention a card only when the host renders it.
  6. On cancellation, call cancel_backtest only for the selected active task and confirm status: cancelled using get_backtest. A pending cancellation is not completion.

If authorization fails, use the host's connect or reconnect action, complete browser authorization, and resume only after the host confirms the connection. Keep the original request. Report returned service failures without inventing results or substituting a local engine.

When saving reports, preserve the returned filename; append a numeric suffix before .zip if it already exists. Historical results do not establish future returns.

Delete a selected result only on an explicit deletion request. Deleting all Research data requires a separate explicit request and the service's confirmation value; it does not delete the user's external identity account.

Related Skills