KB / mcp

MCP tool: read_report

Last verified 2026-05-26

read_report returns the platform’s flagship Markdown research report — the same report.md the engine writes after every cycle and the dashboard renders on the home page. Call this when market_pulse isn’t enough and you need the full narrative across equity markets, macro, volatility, credit, energy, sectors, gamma walls, Fed Watch, correlations, and “what changed” since the last cycle.

Signature

read_report(
  filename: str | None = None,  # specific archive name; omit for latest
) -> str                        # JSON-serialised dict

The tool wraps GET /api/v1/reports/latest when no filename is given, or GET /api/v1/reports/{filename} for a specific archive (report-YYYYMMDD-HHMMSS.md shape).

Returns

{
  "filename": "report-20260526-094500.md",
  "date": "2026-05-26",
  "time": "09:45:00",
  "generated_at": "2026-05-26T09:45:00-04:00",
  "content": "# Market Report — 2026-05-26 09:45 ET\n\n## Market Overview\n...",
  "previous_filename": "report-20260526-080000.md",
  "next_filename": null
}

content is the full Markdown body — sections in fixed order: Market Overview, Sector Breadth, Gamma Call/Put Wall, Fed Watch, Correlations, Macro Fundamentals, Energy, What Changed.
previous_filename / next_filename let the agent walk the archive without a separate listing call. The latest report has next_filename: null.
generated_at is ET ISO 8601 with offset (Law 1).

Behaviour

Cache TTL. CACHE_TTL_REPORT = 30s (mcp/mcp_server/config.py). Reports are written once per cycle (cadence depends on profile — hot every 5 min, market every 15 min during RTH, full 4×/day); a 30s cache deduplicates burst reads cheaply.
Source path. The latest report is /app/data/report.md inside the backend container; archives are report-YYYYMMDD-HHMMSS.md in the same directory. The route reads from disk — no DB round-trip — so this tool stays fast even when the signals DB is hot.
Read-only, no auth. The route is public.
Rendered via Jinja2 with the num filter. Per Law 6, missing numeric inputs render as "N/A" rather than leaking $0.00 or 0.00%. An agent reading "N/A" should treat the field as unavailable, not zero.

Examples

Latest report

Operator: "Read me the latest market report."
Agent: read_report()

Returns the most recent cycle’s full Markdown.

Specific archive

Operator: "What did the report say at the open today?"
Agent: read_report(filename="report-20260526-093100.md")

Walk the archive

Agent: read_report()
  -> previous_filename = "report-20260526-080000.md"
Agent: read_report(filename="report-20260526-080000.md")
  -> previous_filename = "report-20260526-063000.md"

The previous_filename / next_filename chain lets an agent reconstruct a session timeline without a separate listing endpoint.

When to use

Building a thesis — read the full narrative across all sections at once.
Following up on a market_pulse headline that needs more context.
Reviewing prior cycles when the operator asks “what was the report saying at noon?”

When NOT to use

You just need the regime + score + alerts (use market_pulse).
You need structured data, not narrative — use analyze_signals for typed JSON.
You need a specific signal value — go straight to the matching analyze_signals aspect.