AgenticTradingFloor

A multi-LLM trading simulation platform where four autonomous agents analyze real-time market data and make coordinated trading decisions. Built with MCP for seamless tool integration and data-driven experimentation in a realistic trading environment.

Highlights

• Autonomous, multi-agent trading loop with alternating trade/rebalance cycles
• Live market data via Polygon (or free EOD fallback)
• Built-in MCP servers for Accounts, Market, Push; optional Search + Memory servers
• Gradio dashboard with live PnL, charts, holdings, transactions, logs
• SQLite persistence for accounts, logs, and cached market data

Quickstart

• Install uv: see https://docs.astral.sh/uv/
• From the project root:

cd /Volumes/SpaceHD/Learn/AgenticTradingFloor
uv sync

Run the dashboard

• Visualize all traders in a single view:

uv run src/app.py

Run the trading floor (scheduler)

• Runs every N minutes; alternates between trading and rebalancing:

# Optional environment knobs
export RUN_EVERY_N_MINUTES=60
export RUN_EVEN_WHEN_MARKET_IS_CLOSED=false
export USE_MANY_MODELS=false

uv run src/trading_floor.py

Data sources: Live vs Free EOD

• Set Polygon credentials to use live or delayed intraday data; otherwise falls back to EOD or random placeholder prices.

Required for live/deferred intraday:

export POLYGON_API_KEY=your_key
# Options: "paid" (no real-time trades/quotes), "realtime" (full real-time), or leave unset for EOD
export POLYGON_PLAN=realtime

Seamless tool integration (MCP)

Agents use MCP tools that are auto-wired at runtime:

• Accounts: get_balance, get_holdings, buy_shares, sell_shares, change_strategy
• Market (local or Polygon MCP): lookup_share_price (local) or rich Polygon tools (snapshot, last trade, fundamentals)
• Push notifications: push
• Researcher toolbox: web fetch, Brave Search, and persistent memory

These are orchestrated via stdio MCP clients (no manual wiring needed to run the platform).

Manual MCP servers (optional)

Run any server by hand for debugging or external client connections.

• Local Market (EOD or fallback):

uv run src/market_server.py

• Polygon Market MCP (live/delayed intraday):

export POLYGON_API_KEY=your_key
uvx --from git+https://github.com/polygon-io/mcp_polygon@v0.1.0 mcp_polygon

• Accounts:

uv run /Volumes/SpaceHD/Learn/AgenticTradingFloor/src/accounts_server.py

• Push notifications (Pushover):

export PUSHOVER_USER=your_user
export PUSHOVER_TOKEN=your_token
uv run src/push_server.py

Manual “Search + Memory” MCP servers (for research workflows)

• Web fetch:

uvx mcp-server-fetch

• Brave Search:

export BRAVE_API_KEY=your_key
npx -y @modelcontextprotocol/server-brave-search

• Memory (LibSQL, local db per trader):

The app will also auto-start these internally when running the agents; manual startup is only needed if you prefer explicit processes or want to attach external MCP clients.

Environment variables

• Trading/agents:
  • RUN_EVERY_N_MINUTES (default 60)
  • RUN_EVEN_WHEN_MARKET_IS_CLOSED (default false)
  • USE_MANY_MODELS (default false; toggles multiple vendor models)
  • Optional LLM keys: OPENROUTER_API_KEY, DEEPSEEK_API_KEY, GROK_API_KEY, GOOGLE_API_KEY
• Market data:
  • POLYGON_API_KEY
  • POLYGON_PLAN = paid | realtime | unset (EOD)
• Search:
  • BRAVE_API_KEY
• Push:
  • PUSHOVER_USER, PUSHOVER_TOKEN

Reset demo accounts

• Resets four themed strategies and clears balances/holdings history:

uv run src/reset.py

Data storage

• SQLite database: src/accounts.db
  • Tables: accounts, logs, market (cached grouped daily aggs)

Troubleshooting

• No prices or zero quotes:
  • Check POLYGON_API_KEY and POLYGON_PLAN, or expect EOD/free fallback
• Brave Search errors:
  • Rate limited or missing BRAVE_API_KEY; the fetch server can still retrieve web content
• Push not received:
  • Verify PUSHOVER_USER and PUSHOVER_TOKEN
• UI not updating:
  • Keep the scheduler running; the UI polls and reads from the database

What’s inside (at a glance)

• Core loop: /src/trading_floor.py
• Agents and tools: /src/traders.py, /src/templates.py, /src/tracers.py
• Market data: /src/market.py, /src/market_server.py
• Accounts: /src/accounts.py, /src/accounts_server.py
• Push server: /src/push_server.py
• MCP wiring: /src/mcp_params.py
• Dashboard: /src/app.py