Modern research is part data gathering, part synthesis. Analysts still lose hours copying figures from filings, pulling valuation metrics, and checking whether a headline actually moved the tape. The bottleneck is rarely judgment) it is friction between the question in your head and the primary sources that answer it.
Agent-assisted IDEs change that math when they can reach primary sources directly, not scraped summaries or stale PDFs, but the same structured endpoints your production systems use. The analyst stays in the loop; the agent handles the tedious fetches and formatting.
We have published Cursor-friendly workflows that connect agent mode to the Vantafin MCP server at mcp.vantafin.com. The server exposes 23 tools: get_time, search_companies, get_quotes, get_company_profile, get_valuation_metrics, get_financials, get_revenue_segmentation, get_etf_holdings, get_price_history, search_filings, list_filings, read_filing, list_transcripts, read_transcript, search_transcripts, search_earnings_materials, get_earnings, screen_stocks, get_news, get_events, get_macro, get_corporate_actions, and get_ipos. Ask for a company's last four quarters of margin trends, compare megacap retailers on dividend metrics, or draft a memo that cites exact filing language.
Configure the connection
Install the MCP server using the setup guide in our docs. Add your API key to the local config Cursor reads at startup. Once connected, verify the link with a simple prompt such as "fetch the latest 10-Q filing date for MSFT" via list_filings) before you rely on it in a live workflow. A failed auth at 6 a.m. before earnings is an avoidable surprise.
After connection, the agent can call the same data you would reach through the REST API (quotes, financials, screeners, earnings, dividends, and more) without you writing glue code for each new question. The MCP layer handles authentication, pagination, and response shaping so prompts stay focused on analysis rather than HTTP details.
Typical workflows we see in the wild:
- Earnings prep (pull transcript passages with
search_transcriptsand compare guidance language using the workflow in our management guidance guide. - Event triage) when a material 8-K hits, ask the agent to summarize items via
search_filingsand check halt status on your WebSocket monitor. - Peer sets (generate a comp table from
screen_stocksor screener output and save tickers to watchlists.
Document which tools the agent used in your memo footer if your compliance team cares about reproducibility. A one-line citation such as "data via Vantafin MCP, accessed [date]" is usually enough.
Make it repeatable
Ad-hoc prompts are powerful; saved templates are what turn a demo into a process. Save prompts as Cursor rules or project instructions so every coverage initiation follows the same checklist: business overview from get_company_profile, three-year margin stack from get_financials, short interest from the REST API, and open alerts for the next earnings date on the calendar. The agent handles fetches; the template enforces rigor.
Version your templates when you add a new data source. If you start pulling ETF holdings for sector context via get_etf_holdings, add that step explicitly so a junior analyst does not skip it. Repeatability is how you scale quality across a team without everyone reinventing the same Cursor session from scratch.
Run a monthly review of failed or incomplete agent runs. MCP timeouts, rate limits, and ambiguous prompts show up in logs before they show up in a published note. Tuning prompt wording to "return a markdown table with sources" often fixes more than buying faster hardware.
Why we built this
Vantafin was designed as infrastructure analysts can trust. MCP makes that infrastructure conversational without dumbing it down: numbers still come from the same normalized pipelines, and filings still link to accession IDs you can audit. The interface changes; the data contract does not.
Read more about our approach on the about page, or compare access tiers on pricing. If you are evaluating agent tooling for a research desk, start with one coverage name and one workflow. Earnings prep is a good first candidate before you wire MCP into every morning meeting.
Wiring agents to primary sources
IDE agents excel at glue work: fetching data, transforming it, and drafting memos, but only when they reach authoritative sources instead of inventing numbers. Vantafin's MCP server exposes 23 tools on the same normalized pipelines that power the REST API and the app: search_filings, search_transcripts, get_price_history, screen_stocks, get_events, get_ipos, get_macro, get_etf_holdings, and the rest documented in the MCP server guide.
Install the server using the setup block in our docs, add your API key to the config Cursor reads at startup, and verify the link with a simple prompt such as "fetch the latest 10-Q filing date for MSFT") before relying on it in a live workflow. Structured prompts outperform vague asks: request tables, bull/bear bullets, or a checklist tied to SEC filings and transcripts.
MCP shines for ad-hoc exploration; REST shines for deterministic jobs. Many teams prototype in Cursor with get_financials and get_quotes, then promote stable fetches to cron once the query shape stops changing. That handoff keeps research fast without letting experimental prompts run unattended overnight.
Prompt patterns that survive review
Run the same topic query across a screener short list (cloud capex mentions across mega-cap software names, for example) and ask the agent to produce a comparison table with citations. Peer comparison prompts surface divergence faster than reading transcripts sequentially. When the agent returns numeric claims, spot-check against bulk quote snapshots and financials pages; agents are fast, but primary sources remain authoritative.
Combine search_filings with read_filing-style workflows through list_filings plus semantic search when you need exact 8-K language. Link every claim back to Vantafin source pages in the memo footer if compliance cares about reproducibility. A one-line citation is usually enough. Agents can fire many tool calls per turn, so monitor usage headers and cache stable answers within a session rather than re-fetching identical data on every follow-up.
Document which prompts produced reliable memos and which needed correction. A simple log of prompt, tools invoked, and reviewer notes compounds into team knowledge faster than sharing one-off chat exports.
Rate limits, security, and team norms
Compare tier limits on pricing before automating wide-universe sweeps through an agent loop. A prompt that scans five hundred names via screen_stocks can exhaust daily quotas in a single session if nobody watches the counter. Scope agent access to the minimum tool set your workflow needs, especially in shared workspaces where config files travel between machines.
Rotate keys from settings if a config file leaks or a contractor finishes a project. Financial research agents should never be the only control on material non-public information boundaries) human review before trading is non-negotiable regardless of citation quality. For teams already running material 8-K event monitors, agents make excellent second-pass summarizers on email filing alerts that arrived overnight.