FriesTrader

An AI trading agent built to run cheap and fully on its own, trading real orders on Robinhood using its Agentic Trading MCP server. Once set up, it runs unattended on its own schedule every weekday, no manual triggering needed. Two short scheduled Claude Code sessions a day screen stocks, write out their reasoning, and (only under a narrow, explicit gate) place real trades, without a team of specialized sub-agents burning tokens on every decision. The actual safety mechanism is mechanical, auditable risk rules, not the model's judgment, and because it's just two lean sessions instead of a multi-agent pipeline, it runs comfortably on a Claude Pro subscription (as low as $200/year on the annual plan), no Claude Max or metered API spend required.

This is a template/framework extracted from a real, live deployment. Adapt it, don't just run it blind — read "What this does and doesn't solve" below before pointing it at real money.

Why this is safer than it sounds

"Fully autonomous" and "trading real money" together should make you nervous. Here's what actually stands between a thesis and an order:

• Every trade passes through mechanical rules the LLM cannot override — position sizing, stop-loss, take-profit, daily/weekly loss limits, a wash-sale guard. A good story never cancels a stop-loss.
• New deployments start in dry_run and stay there for a minimum number of cycles (dry_run_min_cycles_before_live) before a live order is even possible, so you can watch it screen and reason before it touches real money.
• Only you can flip execution.mode to "live" — the agent is explicitly barred from ever changing this itself, and refuses to place live orders while dry_run.
• Every decision is logged, approved or rejected — trade_log.jsonl is append-only, so you can check whether the reasoning is actually sound, not just trust it.

Requirements

• A Robinhood account with Agentic Trading enabled, connected via Robinhood's own MCP server.
• Claude Code, on a Pro subscription or higher.
• A GitHub account, to host your own fork of this repo.

How it works

Trading runs as two separate phases, on two separate schedules — a full trading day's closing data feeds the thesis, and a fresh opening price is used for the actual order, rather than trading on a stale overnight price.

graph TD
    RH[Robinhood MCP] -- watchlist + scan / quotes / historicals --> A[Phase A: Screen & Thesis]
    A -- thesis per candidate --> P[pending_proposals.jsonl]
    P --> B[Phase B: Re-verify & Risk Enforcement]
    RR[risk_rules.json] -- mechanical limits --> B
    RH -- fresh open price / positions --> B
    B -- dry_run or gated live order --> RH
    B -- every decision logged --> L[trade_log.jsonl]
    L -- plain-English recap --> REC[trade_log_recent.md]

• Phase A (Steps 1–3, ~4:30pm Central weekdays) — screens candidates from your watchlist plus a supplementary market scan, gathers signals, writes a logged thesis per candidate to pending_proposals.jsonl. Places no orders, not even dry-run ones. Full spec: PHASE_A_TASK.md.
• Phase B (Steps 4–7, ~8:35am Central weekdays) — re-verifies Phase A's proposals against fresh opening data, enforces risk_rules.json mechanically, and dry-runs or (gated) places orders. Full spec: PHASE_B_TASK.md.

Both are designed to run as cloud-hosted scheduled agent sessions, independent of any local machine — each run clones this repo fresh and commits/pushes its results back to main, so the repo itself is the persistent state, not local disk.

• risk_rules.json — the hard, mechanical limits (position sizing, stop- loss, loss limits, universe filters, execution mode, wash-sale guard). Nothing in this system should be able to override these. Several fields need your own account details before this is usable — see First-time setup below.
• PHASE_A_TASK.md / PHASE_B_TASK.md — the full, self-contained spec each phase follows.
• trade_log_template.jsonl — the log line shapes; real logs accumulate in trade_log.jsonl in this same style.

See it in action

This is what a real Phase B cycle actually produces (trade_log_recent.md, regenerated every run, symbols genericized):

2026-07-09

Loss limit: OK — daily 0.0%, weekly -2.1%, within -5%/-10% limits.

Held positions (stop-loss / take-profit):

• EXAMPLE — stop 7.00% (vol-scaled), drawdown -2.3% — holding

New-entry candidates considered: OTHER, ANOTHER

• OTHER — approved: medium conviction, $60.00 (12% of account)
• ANOTHER — rejected: max_concurrent_positions already filled this cycle

Orders placed: OTHER — buy $60.00 (dry_run)

No JSON parsing required to see what it did and why. Full field-level examples (thesis records, raw trade_log.jsonl lines) are further down in Example output.

What this does and doesn't solve

• It gives you a structured, auditable version of "let an LLM screen and reason about trades" instead of an opaque one.
• It does not make LLM-driven stock picking more likely to beat a simple index fund — there's no established track record for that, and this can't backtest the reasoning step honestly (news-based reasoning can't be validated against historical data the model may already know the outcome of).
• The risk rules are the actual safety mechanism here, not the reasoning quality. Treat loosening them as the highest-risk change you can make to this system.
• This is a template extracted from a real deployment trading a small personal account, shared for others to learn from or adapt. It is genuinely not financial advice, and running it against real money is entirely your own decision and risk.

First-time setup

1. Fork this repo (or otherwise create your own copy) to your own GitHub account — make it private, since it'll accumulate real trading data (trade_log.jsonl, proposals) once running. Phase A/B commit and push results back to main, so you need a repo you actually control, not this one.
2. Robinhood's Agentic Trading requires a separate, dedicated account — distinct from your regular investing account, and restricted to only the funds you put in it. See that page to open one and connect its MCP server to Claude Code (or to your routine's MCP connections). Nothing below works without this: every tool call in PHASE_A_TASK.md/PHASE_B_TASK.md (quotes, positions, orders, etc.) goes through it.
3. Fill in account_number in risk_rules.json with your own Robinhood account number, set starting_capital_usd to your real starting balance, set universe.watchlist_name to a watchlist you've already created and populated in your Robinhood account, and review every other threshold — the defaults here are illustrative, not a recommendation.
4. Create a scan via the Robinhood MCP's create_scan tool (relative volume > 2.0x, market cap above your min_market_cap_usd), then paste its ID into universe.supplementary_scan_id. Phase A calls this scan every run to surface movers outside your watchlist — left as the placeholder, that call fails every cycle.
5. Fill in wash_sale_avoidance.linked_accounts with every Robinhood account number you personally control, not just this one — if this is genuinely the only account you trade in, a single-entry list (just this account's number) is enough. Leave enabled: true unless you specifically want buys never blocked on wash-sale grounds.
6. Keep execution.mode set to "dry_run". Leave it there for at least the number of cycles set in dry_run_min_cycles_before_live — don't shortcut this.
7. After each cycle, read trade_log.jsonl yourself. Look specifically at rejected candidates and stop-loss triggers, not just the trades that "worked" — that's where you'll see if the reasoning step is actually sound or just getting lucky with an uptrend.
8. Only flip execution.mode to "live" yourself, by hand, after you've reviewed enough dry-run cycles to trust the output. Do not let the agent flip it for you as a shortcut.

Keeping your fork updated

This template gets improvements over time.

• Easiest: GitHub's "Sync fork" button, on your fork's main page. No local git needed. Works cleanly as long as nothing conflicts with your own changes.
• If that button refuses (conflicts, usually in risk_rules.json), resolve locally:

  git remote add upstream https://github.com/YizhiSong/FriesTrader.git
  git fetch upstream
  git merge upstream/main
  

  Resolve any conflicts in risk_rules.json by hand — your own account details and thresholds should win, not upstream's placeholders.

Running it

Two schedules need to fire: Phase A around 4:30pm Central on weekdays (hand Claude Code PHASE_A_TASK.md to execute), and Phase B around 8:35am Central on weekdays, 5 minutes after market open (hand it PHASE_B_TASK.md). Each run is a fresh Claude Code session pointed at this repo — no state needs to persist locally between runs, since the repo itself (risk_rules.json, pending_proposals.jsonl, trade_log.jsonl) is what's read and written each time.

• Recommended: Claude Code's own scheduled cloud routines. Set one routine to run PHASE_A_TASK.md on the Phase A schedule and a second for PHASE_B_TASK.md on the Phase B schedule, with the routine's source pointed at your fork from First-time setup, not this repo. This runs independent of any machine being on — the actual point of