ask chatgpt to find you founders building in stealth and you will get a decent essay about linkedin boolean searches.
that is not a model problem. it is a reach problem. a stealth founder is, by definition, someone with almost no public web presence. the model is excellent at reasoning and has nothing to reason over.
so the fix is not a better prompt. it is giving the agent a source that sees things the public web does not.
what the agent is missing
three kinds of signal sit outside what a general model can retrieve:
who started following whom, and when. a follow-graph diff only exists if something was recording on the day it happened. it is not on a page anywhere for a crawler to find, and it cannot be reconstructed later by anyone, including us.
a judged verdict. raw signal is mostly noise. of 2,478 people flagged building in stealth, 896 clear a career check. an agent handed the unfiltered 2,478 produces a worse answer than one handed the 896, because it will confidently rank the noise.
a date that predates the announcement. models are good at describing what happened. the useful question in sourcing is what has not happened yet.
the setup
in claude code or cursor:
npx frontrun-mcp-server --setup
that installs 37 tools over mcp. the oauth wall appears at first use, which is where a trial starts. for a remote connection instead, point your client at https://www.frontrun.vc/mcp with a bearer token from settings.
mcp and api access are on the pro plan at $99/mo.
the first prompt worth running
once connected, the useful thing is not a search. it is a chain the agent runs on its own:
show me people flagged building in stealth in the last 14 days with a strong verdict. for each one, pull the evidence behind the verdict and their previous employers. drop anyone whose background is not relevant to applied ai or robotics. rank what is left by how many tracked investors are already following them, and tell me which single one you would reach out to first and why.
every step there is a separate tool call, and the agent decides the order. it surfaces the lane, filters on verdict, pulls evidence, applies your thesis, and ranks on investor attention. what you get back is a short list with reasons, not a csv.
the follow-up that matters:
for the top one, draft a first message. do not mention that they are in stealth, do not name the company, and lead with something specific to their previous work.
why "do not mention stealth" is in the prompt
because the most common way to waste this signal is to open with it.
a founder who has told nobody does not want an email that proves a stranger is watching. the signal tells you who to talk to and when. it should not appear in the message itself. an agent will happily write "i saw you recently went into stealth" unless you tell it not to.
what this is not good at
worth stating plainly, because an agent will not tell you:
- it does not know if the idea is good. it can tell you a credible person started building and that investors noticed. whether the thing works is still your judgement.
- it cannot go backwards. ask for who was following someone last march and there is no answer. the diff was either recorded or it was not.
- it is not a company database. if you want exhaustive coverage of every incorporation, use one of those. this is the short list, not the index.
the shape of the result
the point of wiring this up is not automation for its own sake. it is that the loop from "something moved" to "here is the person, here is why they are credible, here is a draft" collapses into one prompt, and it runs against signal your competitor's agent cannot query.
the receipt for whether that matters: jacob zietek was published on our public stealth lineup on august 19 2026. twenty-one days later he announced a pre-seed from a16z and accel. an agent pointed at this data on august 19 would have surfaced him. an agent pointed at the public web would have found nothing, because there was nothing.
related
- how to find founders building in stealth, the nine signals, agent or no agent
- the ixian robotics timeline, the dated receipt in full
- agentic deal flow, what the agent stack looks like end to end
- ai deal sourcing
open source deal flow
companies and founders surfaced on frontrun. curated, every week.
