what if you didn't have to doomscroll x to find startups?
that's the whole pitch. frontrun watches the follow graphs of 2,000+ tracked venture investors on x, around the clock, and flags the companies they start paying attention to before the news exists. the mcp (model context protocol) server puts that entire signal layer directly inside claude. you type a sentence, your agent does the sourcing.
What if you didn't have to doomscroll @X to find startups?
— frontrun (@frontrunvc) August 5, 2026
Introducing the @frontrunvc MCP.
Claude reads 2k+ accounts across AI, Crypto, and Tech and surfaces companies and founders early.
One prompt: 10 robotics companies under 1k followers, every founder, in seconds. pic.twitter.com/pOUcuY021H
this guide is the practical setup: claude on the web, claude desktop, claude code in the terminal, and your own agents (hermes, openclaw, cursor, anything that speaks mcp). if you want the conceptual overview of mcp for investors first, read mcp for venture capital.
what "agentic deal flow" actually means
traditional deal sourcing tools give you a database and a search box. you do the work. agentic deal flow flips it: your ai agent holds the tools, and you hold the thesis.
connected to frontrun's mcp, claude gets 35 read-only tools over the live investor follow graph: trending companies, convergence events (2+ tracked investors following the same startup), semantic thesis search, founder resolution, funding history, and first-flagged provenance for every company. the data is the part you can't rebuild with a scraper: continuous follow-graph diffs across 2,000+ investors, running for months, with receipts.
the result is a startup discovery workflow that runs at the speed of a prompt. no tabs, no csv exports, no doomscrolling.
connect frontrun directly in claude (60 seconds, no code)
the fastest path is the custom connector on claude.ai and claude desktop. no install, no key handling:
- open claude → settings → connectors → add custom connector
- paste the server url:
https://mcp.frontrun.vc
- claude opens a frontrun login window. sign in, click connect, done.
behind the scenes this is oauth: claude gets your personal mcp key, revocable anytime at frontrun.vc → settings → api keys. mcp access is included with frontrun pro ($99/mo, 10,000 monthly credits).
connect claude code (two commands)
for the terminal and ide crowd, the server is on npm as frontrun-mcp-server:
npx frontrun-mcp-server --login # opens browser, oauth sign-in
npx frontrun-mcp-server --setup-code # registers the server with claude code
--setup does the same for claude desktop, and cursor users can point at the same package. full client-by-client instructions live in the mcp server docs.
what your agent can do once it's connected
| you type | your agent does |
|---|---|
| "what are tracked investors converging on this week?" | ranks companies by net-new tracked follows, with flag dates |
| "find agent infrastructure startups under 1,000 followers" | semantic thesis search + live follower filters |
| "who founded @scalar_field_ and who's backed them?" | founder resolution + funding history in one answer |
| "which of these have 2+ tier-1 funds following?" | convergence check across the follow graph |
| "draft intros to the top 5 in my voice" | combines the signal with your context and writes the outreach |
every answer carries provenance: when we first flagged the company, at what follower count, and which tracked investors followed since. that's the difference between an ai agent guessing and an ai agent citing.
receipts: a published article, sourced in one sentence
this isn't a demo on staging data. 10 companies helping you own your own model, published on this blog and posted to x, was sourced with a single sentence in claude code connected to frontrun's mcp: thesis in, companies + founders + funding out.
here's the full session, end to end:
using claude code + @frontrunvc mcp to source the companies for this article
— JG (@jongall45) July 30, 2026
thesis to companies, founders, and fundraise info in a few prompts. https://t.co/G9bzUZKaNt pic.twitter.com/cboKO6PdoM
three of the ten companies that came back, straight from the follow graph:
Hyde: training + inference for models enterprises own. "frontier models are for renting. your intelligence isn't."
Thirdbrain Labs (a16z speedrun): the post-training layer. data in, your model out.
distil labs: a custom small language model from a prompt, via distillation. deploy it as a drop-in llm replacement, own it outright.
sub-1k-follower companies, resolved founders, verified rounds. one prompt each to go deeper on any of them.
beyond claude: hermes, openclaw, and custom agents
mcp is a standard, so the same server plugs into whatever agent stack you run:
- hermes, frontrun's hosted deal flow agent, ships with the signal layer built in: a 24/7 agent that watches the follow graph for you and messages you when something moves.
- openclaw agents take the same mcp config block and get the full toolset: tracking, convergence detection, classification, search. setup in the openclaw docs.
- cursor, codex, and any mcp client work with the same two commands above.
- building your own harness? skip mcp and hit the rest api directly: 24 endpoints under
/v1, same auth, same credits.
faq
do i need to code? no. the claude.ai connector is paste-a-url. the terminal setup is two copy-paste commands.
which clients are supported? claude (web + desktop), claude code, cursor, codex, openclaw, and anything that speaks mcp.
what does it cost? frontrun pro, $99/mo: web app + mcp + api + 10,000 monthly credits. the $49/mo starter covers the web app only.
can i try the data first? the trending page is a free public slice, no login.
is it safe to connect? the mcp tools are read-only over your account's data, auth is oauth, and you can revoke the key anytime from settings.
follow from the front. connect your agent: frontrun.vc
read next: mcp for venture capital · how to build an ai deal flow agent · best ai deal sourcing tools in 2026
