The MCP server is a thin, tenant-scoped wrapper over /api/agent/v1. It exposes research universes and factors, market data, chart-indicator authoring, Strategy API V2 deployment and backtesting, broker/execution observations, notification automation, bounded jobs, and safety-gated trading.
pip install "quantdinger-mcp==0.5.0"
export QUANTDINGER_BASE_URL=http://localhost:8888
export QUANTDINGER_AGENT_TOKEN=qd_agent_xxx
quantdinger-mcppipx install quantdinger-mcp and uvx quantdinger-mcp are also supported. Use pip install -e ./mcp_server only when developing from a repository checkout.
The default transport is stdio. Set QUANTDINGER_MCP_TRANSPORT to sse or streamable-http for a network transport. Optional limits include QUANTDINGER_TIMEOUT_S, QUANTDINGER_MCP_JOB_STREAM_MAX_EVENTS, QUANTDINGER_MCP_JOB_STREAM_MAX_SECONDS, and QUANTDINGER_MCP_JOB_POLL_MAX_SECONDS.
Network transports bound to a non-loopback host require a separate inbound bearer token. This token authenticates MCP clients and must not be the Agent Gateway token:
export QUANTDINGER_MCP_TRANSPORT=streamable-http
export QUANTDINGER_MCP_HOST=0.0.0.0
export QUANTDINGER_MCP_PORT=7800
export QUANTDINGER_MCP_PUBLIC_URL=https://mcp.example.com
export QUANTDINGER_MCP_AUTH_TOKEN=replace-with-a-random-32-plus-character-secret
quantdinger-mcpClients must send Authorization: Bearer <QUANTDINGER_MCP_AUTH_TOKEN> to /mcp or /sse. Authenticated non-loopback listeners require an HTTPS QUANTDINGER_MCP_PUBLIC_URL. QUANTDINGER_MCP_ALLOW_HTTP=true is only for a trusted private proxy that terminates TLS. For an unauthenticated private ingress that already authenticates every request, QUANTDINGER_MCP_ALLOW_INSECURE_HTTP=true remains a separate escape hatch; never use either setting on a directly reachable public listener.
Docker builds use the official PyPI index by default. In regions where it is slow, override it without editing the image definition: docker build --build-arg PIP_INDEX_URL=https://your-mirror.example/simple ..
Never place an agent token in prompts, logs, screenshots, source control, or MCP configuration that will be shared. Responses redact credential fields, and clients must not attempt to recover them.
| Tool group | Scope | Purpose |
|---|---|---|
whoami, check_health |
R/public | Identity, allowlists, and liveness |
list_markets, search_symbols, get_klines, get_price |
R | Market discovery and data |
| Universe and factor tools | R | Point-in-time research inputs |
list_watchlist, add_watchlist, remove_watchlist |
R/W | Watchlist workspace |
| Indicator authoring, validation, save, link, and read tools | R/W | Chart-only indicators |
list_strategy_templates, compile_strategy_code |
R | Strategy API V2 templates and manifest compilation |
list_strategy_sources, get_strategy_source, save_strategy_source |
R/W | Private Strategy API V2 source library |
list_strategy_source_versions, restore_strategy_source_version |
R/W | Source history and explicitly confirmed restore |
create_strategy, update_strategy, list_strategies, get_strategy |
R/W | Strategy API V2 deployments |
submit_backtest |
B | Strategy API V2 backtest job |
list_jobs, get_job, wait_for_job, stream_job_until_done, cancel_job |
R/B | Bounded jobs and confirmed cancellation |
runtime_overview, stop_strategy |
R/T | Runtime inspection and confirmed stop |
| Broker account, strategy position/trade, and quick-trade observation tools | R | Secret-free execution observations |
| Signal-alert tools | N | Notification task lifecycle and confirmed delivery evaluation |
place_quick_order |
T | Confirmed order with token notional caps |
list_portfolio_positions, list_paper_orders |
R | Portfolio and paper-order reads |
emergency_stop_trading, cancel_open_paper_orders |
T | Emergency cancellation and T-token revocation |
Every mutating W/B/N/T tool requires a caller-generated idempotency_key; retries of the same request must reuse it. stop_strategy requires confirm_stop=true. place_quick_order requires confirm_order=true; a live-capable token also requires confirm_live_trading=true. Optional tp_price and sl_price protection are forwarded to the shared Quick Trade execution path. Server-side trading flags, allowlists, and per-order/per-day notional caps still apply.
Executable strategy code must define initialize(context) and declare its universe and subscriptions. It must provide handle_data, on_rebalance, or a scheduled callback. The manifest owns instruments, markets, frequencies, factor dependencies, warmup, and leverage policy.
Compile and save a source before creating a stopped deployment:
compile_strategy_code(code="...Strategy API V2 Python...")
save_strategy_source(name="btc-momentum", code="...Strategy API V2 Python...")
Use the returned source id:
create_strategy(
name="btc-momentum",
source_id=12,
initial_capital=10000,
execution_mode="signal",
params={"lookback": 40},
idempotency_key="deploy-btc-momentum-v1"
)
Run a backtest directly from V2 code:
submit_backtest(
code="...Strategy API V2 Python...",
start_date="2025-01-01",
end_date="2025-12-31",
initial_capital=10000,
params={"lookback": 40},
idempotency_key="btc-momentum-2025"
)
Market, symbol, and timeframe are not backtest parameters. They come from the compiled strategy manifest. Use wait_for_job or stream_job_until_done to obtain the result.
Indicators are chart-only. Validate and save them through the indicator tools, then convert the idea into Strategy API V2 code before using submit_backtest or create_strategy.
Restoring a source snapshot requires confirm_restore=true. The emergency stop requires confirmation, attempts to cancel agent-originated live orders, cancels paper orders, revokes every active tenant T token, and reports exchange cancellations needing human follow-up.
The optional Docker network service is enabled explicitly:
QUANTDINGER_AGENT_TOKEN=qd_agent_xxx \
QUANTDINGER_MCP_AUTH_TOKEN="$(openssl rand -hex 32)" \
docker compose --profile mcp up -d --build mcppip install -e './mcp_server[dev]'
pytest mcp_server/tests