Leave a record the next session can use
Keep the research question, primary source links, selected universe, rejected candidates and accepted backtest IDs in the Bot workspace. Ask it to check current sources when resuming the task. A remembered thesis may be stale even when the files persist.
A persistent agent with a real workspace
Grok Bot is the agent product with a persistent cloud computer, browser, filesystem and terminal. That workspace can retain a research brief and strategy artifacts across tasks. Its ability to use a terminal makes the Python SDK a practical companion to the account tool connection.
Use a named research assignment with explicit outputs: source notes, a frozen universe, the complete strategy configuration and the returned operation IDs. Keep account access and approval choices deliberate.
The tool sequence
Inspect the discovered inputs before submitting a call. For a new candidate, validate and save it first; for a saved candidate, start with get_portfolio. Record the portfolio and backtest identifiers returned by the account tools.
| Validate a new candidate | build_portfolio | Canonical configuration and component issues; nothing is saved |
| Save the validated draft | create_portfolio | Use the reviewed configuration and retain the returned chat portfolio ID |
| Read the saved book | get_portfolio | Exact strategies, action sizing, and portfolio identity |
| Submit the historical test | backtest_portfolio | Explicit dates, interval, capital and benchmark |
| Poll the same operation | query_backtest_status | Terminal state and warnings; do not resubmit while running |
| Inspect the curve | query_backtest_history | Historical values for the returned backtest_id |
| Compare candidates | compare_backtests | Identical period and assumptions, with activity and risk |
No matching rows. Clear the filter to see all records.
Read the actual Grok Bot experiment
Austin’s published case study documents Grok Bot using NexusTrade MCP and the Python SDK to build biotech and semiconductor candidates, followed by a human-directed redesign and manual order approval.
The article separates historical test windows, original paper reproductions and the funded-account implementation. Use it as a workflow example. Its historical tests do not measure the future success of the news thesis, and this page does not restate old performance totals as current results.
Connect to the current NexusTrade surface
Open Marketplace in Grok Bot and add the NexusTrade connector if it is available. Complete browser authentication, then attach it with @ in the task. For an existing or custom MCP connection, ask the Bot to use the endpoint below with OAuth discovery. Open Agent Computer and take control for sign-in or verification, then return control. Call fetch_portfolios with include_live=false and include_positions=false; pass one returned portfolioId as portfolio_id to get_portfolio and compare portfolio.portfolioId, portfolio.type and portfolio.strategies with NexusTrade. An authenticated empty list is valid for a new account.
If the available connector cannot complete that flow, use the documented Python SDK in the Bot’s terminal with a registered API key configured through its secret environment. SDK credentials and MCP OAuth are separate. The case study used an earlier setup; follow the current developer instructions for credentials and available tools.
https://nexustrade.io/api/mcpHand the Bot a research protocol
State the thesis, instrument universe, dates, baseline, fees and maximum candidate budget before testing. Require a frozen historical test window and source availability dates. Ask the Bot to report rejected candidates rather than hiding the selection process.
Research this thesis using dated primary sources. Freeze the candidate universe and write explicit entry, sizing and exit rules. Show the experiment plan before running compute. Keep historical validation and future paper observation separate, preserve every accepted operation ID, and stop before deployment or brokerage approval.Review before deployment
The case study’s practical lesson is that attractive historical totals did not settle the allocation design. Inspect the portfolio’s actual configuration, exposure and exits before accepting a candidate.
Use a new paper deployment to observe forward simulated behavior. Existing brokerage portfolios and any staged orders require their own explicit review. A persistent workspace does not supply approval for brokerage actions.
Choose the task you want to finish
Use the research guide for a dated company brief, the backtest guide for one fixed historical candidate, or the paper guide to observe a separately identified simulated book. Each guide shows how to check the connection, retain the accepted operation ID and recover from a failed request.
Review a brokerage deployment separately
Public and Tradier have native NexusTrade account connection and strategy-port flows described below. Public also offers its own external MCP route. The Robinhood guide uses Claude Code with separate NexusTrade and Robinhood MCP servers; it does not imply that this client already has a verified native Robinhood connection. Broker account selection, order approval and a verified stop remain separate from paper observation.