The useful version of an AI market brief is boring on purpose. It asks the same few questions every morning, it attaches a source to every claim, and every line can be checked in one click. The exciting version — an agent that freely narrates the market — is exactly the one that invents numbers when the data runs out. This post is the boring build: a daily brief agent wired to real market data over MCP.

What a good daily brief contains

Three things, at most: the strongest new signals on your watchlist, notable insider activity, and any filings that fired overnight. Just as important is what it should refuse to contain: predictions, price targets, unsourced summaries, and anything the agent cannot point to a filing or article for. A brief is a reading aid. Its job is to tell you what happened and where to verify it, not what to do about it.

Step 1: connect an MCP client to a market data server

If you have already read how to give an AI agent access to market data with MCP, this is the same foundation: an MCP server exposes read-only tools, your client discovers and calls them. QuantConomy’s hosted server speaks Streamable HTTP at mcp.quantconomy.com/mcp and works with Claude Code, Cursor, or any MCP-compatible client — the connect configs are on the MCP page. One honest caveat: this ships with the beta, and the installable MCP package and tool scopes are still on the finalizing list. Treat this as the workflow to set up, not something to run in production this afternoon.

Step 2: pick the three questions

Resist the urge to ask the agent for “a market overview.” Give it three fixed queries instead:

  • Strongest watchlist signals: list_signals, filtered by your assets and a minimum strength, so only rows that scored high enough make the brief.
  • Insider activity: sec_insider_trades for the same tickers, which returns Form 4 data fresh to the prior day.
  • Filings that fired: sec_current_report_items for overnight 8-K items, plus search_entries if you also want the news coverage around them.

Fixed questions make the output comparable day to day. When the brief looks different, it is because the market did something, not because the model felt chatty.

Step 3: prompt for citations, not summaries

The prompt matters less than one rule inside it: every line in the brief must carry the source link from the tool result, and anything without a tool result behind it gets dropped. Signals from list_signals already arrive source-linked — type, direction, strength, reason, and the filing or article they came from — so the agent’s job is transport, not authorship. A useful test: if you delete the model and read the raw tool output, the brief should say almost the same thing, just less readably.

Step 4: schedule it and spot-check one source per day

Run it on a schedule — a cron job, a scheduled agent, whatever your stack offers — and build one habit on top: open one cited source per day and confirm the brief matches it. That single check keeps the whole setup honest. If a citation ever fails to back its line, tighten the prompt before you trust another brief. An alert you cannot verify is just a rumor with better formatting, and the same goes for a morning brief.

The honest limits

An agent brief is a reading aid, not a strategy. It will not tell you what to buy, and nothing in this setup should — none of this is a recommendation to buy or sell anything. It is also only as good as the questions you fixed in step 2; events outside your watchlist and signal filters will not appear. And if you would rather not run your own, a hosted version is coming: the Signal brief beta is the “Up next” item on our roadmap, a daily read built from the strongest news, SEC, and prediction-market rows.

Start small either way. Two sourced answers every morning beat ten unsourced ones.