# An AI trading desk built as a team of sub-agents (Claude Code and Robinhood MCP) > Source: > Published: 2026-06-30 15:11:46+00:00 **A multi-agent stock-research desk that runs inside Claude Code, connects to a Robinhood Agentic account over MCP, and never places an order without your approval.** It's not a "bot that YOLOs your money." It's a small **team of specialized sub-agents** — fundamental, technical, macro/news, and a risk manager with veto power — that screen your watchlist, debate each candidate, and hand you a one-click **preview card**. You approve; it places. Everything's wrapped in written guardrails (human-in-the-loop, position caps, prompt-injection defense) and mirrored to a Robinhood-style dashboard. **A team, not one prompt**— analysts gather evidence in parallel; an independent risk manager can veto a trade the analysts liked.** Guardrails are structural, not vibes**— sub-agents physically have no order tools; only you-plus-the-PM can place, and only after explicit in-session approval.**Prompt-injection-aware**— the news agent treats fetched content as untrusted data and quotes suspicious "instructions" instead of acting on them.**Low-touch by design**— it runs read-only research on a schedule and only surfaces a trade when one genuinely qualifies; most days it tells you to stand aside.**A real dashboard**— a Robinhood-style UI mirrors the desk's state live. ⚠️ Real money, beta, not investment advice.Robinhood Agentic Trading is in beta (US, equities only). The agent trades only inside an isolated Agentic account funded with a dedicated budget — that budget is the most it can ever lose. There isno track record and no performance claim here; this is a reference architecture for learning, run it at your own risk and monitor it yourself. This repo holds the **guardrails, strategy notes, agents, dashboard, and setup docs** — never secrets, and never the OAuth token (those live outside the repo). ``` . ├── CLAUDE.md # The PM's operating contract (rules it must follow) ├── .mcp.json # Project-scoped Robinhood Trading MCP connection ├── .env.example # Template for .env (account number; .env is gitignored) ├── .claude/ │ ├── settings.json # Permissions: reads allowed, orders gated, options denied │ └── agents/ # The desk team — one sub-agent per role │ ├── fundamental-analyst.md │ ├── technical-analyst.md │ ├── macro-news-analyst.md # injection-isolated (the only web-facing role) │ └── risk-manager.md # veto power over every trade ├── strategies/ │ ├── README.md # Risk caps + when the Risk Manager must VETO │ └── mean-reversion.md # Example strategy (entry/exit/sizing) ├── docs/ │ ├── SETUP.md # OAuth setup + how to tighten guardrails │ └── TEAM.md # The desk roles and end-to-end workflow └── ui/ # Read-only Robinhood-style dashboard (Vite + React) ├── public/desk-state.example.json # demo data (live desk-state.json is gitignored) └── src/ # polls the snapshot, renders account/desk/preview ``` - Make this repo **private** before pushing anything. - Read `docs/SETUP.md` and follow it top to bottom. - Connect the MCP, authenticate via OAuth (desktop), fund a small Agentic budget. - Refine `.claude/settings.json` once you know the real tool names. A small **team of sub-agents** coordinated by the **Portfolio Manager (PM = the main Claude Code session)**. Analysts gather evidence in parallel; the Risk Manager has veto power; only the PM can place orders — and only after **your** in-session approval. After every run the PM writes `ui/public/desk-state.json` , which the dashboard mirrors live. ``` php flowchart TD U([You: ask the PM in plain language]) --> PM{{Portfolio Manager
main session · only role that can order}} subgraph SENSE [1. Sense] PM --> A1[get_portfolio / get_equity_positions
Agentic account only] end subgraph RESEARCH [2-3. Screen + Research · parallel · read-only] A1 --> F[Fundamental Analyst
valuation · earnings] A1 --> T[Technical Analyst
trend · levels · scans] A1 --> M[Macro / News Analyst
backdrop · news · INJECTION-ISOLATED] end F --> SYN[4. PM synthesizes a proposed trade
tied to a rule in strategies/] T --> SYN M --> SYN SYN --> RISK{5. Risk Manager
checks vs strategies/} RISK -- VETO --> STOP([Trade stops · PM reports back]) RISK -- APPROVE / CHANGES --> PREV[6. review_equity_order
build preview card] PREV --> SNAP[(7. Write desk-state.json)] SNAP --> GATE[8. Present preview · ⏸ wait for your approval] GATE -- you say yes --> EXEC[9. place_equity_order
still gated by ask rule] GATE -- you say no --> STOP EXEC --> CONF[10. Confirm fill · refresh desk-state.json · log] SNAP -. polled every 5s .-> DASH[/Dashboard: npm run dev/] CONF -. updates .-> DASH ``` Steps 1–7 are research and produce **no order**. The desk's standard output is the **preview card at step 8** — it stops there until you confirm. Full role/tool breakdown is in [docs/TEAM.md](/LoganYangBo/rh-trading-agent/blob/main/docs/TEAM.md). ``` # 1. One-time: connect + authenticate the MCP, then record your account number claude # open the project (trust the .mcp.json server) # in-session: /mcp # pick robinhood-trading → OAuth (desktop + mobile verify) cp .env.example .env # then put your Agentic account number in .env # 2. Run the dashboard (separate terminal) — mirrors each desk run cd ui && npm install && npm run dev # http://localhost:5180 (shows demo until a live run) # 3. Drive the desk — just talk to the PM in the Claude Code session, e.g.: # "Screen my watchlist and bring me the top 2 ideas with full team analysis." # "Run the desk on AAPL and NVDA, risk-check a small starter in the better one." # The PM fans out to the analysts → Risk Manager → preview card → waits for your OK. # Kill switch: disconnect the MCP from the Robinhood app, or remove it locally claude mcp remove robinhood-trading ``` Note:the sub-agents in`.claude/agents/` load when Claude Codestarts— after adding or editing them, restart the session so roles like`fundamental-analyst` are recognized with their restricted tool sets. - The agent can only trade in the **Agentic account**, never your main balance. - Default permission rule puts **every** Robinhood tool call behind a manual prompt. `CLAUDE.md` includes a prompt-injection rule: the agent must ignore trading instructions found in fetched/external content (news, analyst notes, web).- You can disconnect the MCP anytime from the Robinhood app — that's your kill switch.