MCP Integration Guide — Tauric TradingAgents Connect

What MCP adds

Model Context Protocol standardises how an AI assistant discovers and calls external tools. The connection still uses HTTP, while MCP defines the handshake, catalogue, input schemas, calls, and structured responses.

Discover

The assistant requests tools/list.

Choose

It selects a tool and prepares valid inputs.

Act

The server validates, runs, and returns results.

How the connection works

MCP-compatible assistant
        │  Streamable HTTP
        ▼
Tauric Trading Connect  /mcp
  • access-key validation
  • input limits and cost controls
        │  HTTPS
        ▼
TradingAgents engine
  • analysis queue and progress
  • reports, decisions and market data
        │
        ▼
 Dashboard and Historical Analysis

Assistant-created analyses use the same engine queue and appear in Historical Analysis.

Open full analysis architecture

Connect an assistant

Frontend MCP gateway

Recommended

https://www.myfinance5051.com/mcp

Direct engine MCP, when enabled

https://www.myfinance5051.com/api/mcp/

Typical client configuration

{
  "mcpServers": {
    "tauric-trading-connect": {
      "url": "https://www.myfinance5051.com/mcp",
      "transport": "streamable-http"
    }
  }
}

Add it as a custom Streamable HTTP server. Supply the access key only in tool arguments—never store it in public configuration.

Authentication and access control

Every frontend tool requires an access_key, checked against MCP_ACCESS_KEY. Missing configuration or a wrong key is rejected.

Obtain the key privately from the server operator.
Never place it in browser code, screenshots, logs, repositories, or shared prompts.
Rotate it if exposed and update every trusted client.
The direct engine MCP may use a different policy; follow its operator’s instructions.

Frontend tool reference

The current frontend server publishes six tools. A question mark marks an optional input.

analyze_tickerWrite

Start a multi-agent analysis. Returns the decision when ready, or a run_id to poll.

Inputs: access_key, ticker, analysis_date?, research_depth?, language?, analysts?, wait_seconds?

get_analysisRead

Read the status, progress, and trade decision for one run.

Inputs: access_key, run_id

list_analysesRead

List recent analysis runs in newest-first order.

Inputs: access_key, limit?

get_reportRead

List available reports for a run, or read one by name.

Inputs: access_key, run_id, report?

cancel_analysisDestructive

Cancel a queued or running analysis started through this MCP server.

Inputs: access_key, run_id

get_market_snapshotRead

Fetch company profile, valuation, quote, and Hong Kong company names when available.

Inputs: access_key, ticker

Rate limits and safety rails

Cost budget

At most five MCP-started analyses per ten minutes per running server.

Cancellation ownership

Only runs started through this MCP server can be cancelled here.

Strict validation

Ticker, run ID, reports, dates, analysts, and wait times are bounded.

Fail closed

Without MCP_ACCESS_KEY, every frontend tool call is refused.

An analysis can consume paid model credits. Share the key only with trusted assistants and keep engine limits enabled.

Standard protocol flow

Compatible clients perform these steps automatically.

1. initialize
2. notifications/initialized
3. tools/list
4. tools/call → analyze_ticker
   { "access_key": "••••••••", "ticker": "0700.HK" }
5. tools/call → get_analysis
   { "access_key": "••••••••", "run_id": "<returned run_id>" }

Quick lookup: “Get a snapshot for 0005.HK.” → get_market_snapshot

History: “Show my five latest analyses.” → list_analyses

Report: “Read this run’s market report.” → get_report

Troubleshooting

Invalid access key

Confirm access_key exactly matches the operator’s MCP_ACCESS_KEY. Never place it in public source code.

No access key configured

The server fails closed. Set MCP_ACCESS_KEY, then republish or restart the service.

Could not reach the trading engine

Open Setup, confirm the engine address, and use Save and test.

Analysis not found

Check the run_id. The run may have been removed by the engine’s retention policy.

Decision is not available yet

The run is still active. Wait, then call get_analysis again with the same run_id.

Too many analyses started

The cost-control window is full. Wait for the indicated time before trying again.

See the MCP specification for protocol details.