The research desk
Assemble an AI research desk with MCP data sources, scheduled briefings and searchable notes, while people retain judgment and consequential decisions.
Scenario
A market analyst's day is a loop: pull fresh data, compare it against yesterday's picture, write up what changed and why it matters. The loop is mechanical; the judgment isn't. This playbook automates the loop — live data over MCP, scheduled briefings, durable notes — and keeps the judgment (and every consequential action) human.
The worked example is crypto, because the on-ramp is unusually clean: CoinGecko ships an official, public MCP server (no API key required) covering prices, market caps, volumes, and historical charts for 15,000+ coins. Attach it to an agent and the desk has institutional-grade market data on its first day. The same shape works for any domain with a data API — equities, real estate, commodities, your own industry's numbers.
The team
| Agent | Job |
|---|---|
| Market analyst | The desk's core: reads live data over MCP, maintains watchlists, writes the briefings |
| News scanner | Sweeps the web tools for headlines, filings, and sentiment around the watchlist |
| Archivist | Keeps the desk's memory tidy: daily snapshots into dated files, so "what did this look like in March" is one search away |
The wiring
- CoinGecko's MCP server attaches on the analyst's settings — its tools appear alongside the built-ins, no code written. Other data sources follow the same pattern: an MCP server where one exists, otherwise a skill whose script calls the API with a scoped credential.
- The desk's memory is a synced source: watchlists, briefings, snapshots, theses. Because it is vector-indexed, the analyst's "compare against the last time this happened" is a semantic search, not a guess.
- Charts and dashboards that only exist as web pages are the machine desktop's job — screenshots land in the conversation next to the numbers.
The rhythm
- 07:30 daily — Analyst: morning briefing. Overnight moves on the watchlist,
three things worth attention, one page, into
data://briefs/and the morning conversation. - Hourly — News scanner: sweep, append anything material to the day's log.
- 20:00 daily — Archivist: snapshot the day's data and briefing into the dated archive.
- Weekly — Analyst: the deep dive — one watchlist asset re-examined from a fresh conversation against the archive.
Every run records a result summary on the Calendar, so a glance shows whether the morning briefing actually said anything.
Governance — read this one
A research desk is where governance stops being theory:
- Neuralis does not place trades, and this playbook never wires it to.
The desk produces analysis; execution stays with a human at their own
broker or exchange. If you ever attach a capability that can move money,
that is an
Auto-profile-never decision:Safeprofile, explicit approval on every call, agent-scoped credentials, and a daily cap. - Briefings are research notes, not financial advice — say so on the page the desk publishes to, and read it that way yourself.
- The audit trail is the desk's compliance file: every data pull, every brief, every approval, attributable.
Extend it
Swap the data plane and the same desk covers any market — the MCP catalogs list servers for equities, news, and economic data, and connect anything covers the on-ramps. Point the news scanner at your competitors instead of coins and the desk becomes competitive intelligence for the marketing team.