a short setup guide. five minutes, most of which is the restart.
1. install
npx frontrun-mcp-server --setup
the command asks which client you are configuring and writes the mcp config for it. claude code, cursor and codex all work the same way.
2. restart the client
the most common reason this appears not to work. mcp config is read at startup, so a running claude code will not see the new server until it restarts.
3. verify the tools loaded
do not assume. ask the agent:
list your available tools
you should see the frontrun tools alongside its built-ins. if they are not there, the config went somewhere else, which is the second most common failure: the setup command wrote a config for a client you are not currently using.
4. trigger authentication
the first tool call opens the oauth wall. run something cheap:
use frontrun_status to tell me how many accounts I am tracking
sign in, and the tool call completes. that sign-in is also where a trial starts if you do not have a plan.
5. the remote alternative
if you would rather not install anything, point an mcp-capable client at:
https://www.frontrun.vc/mcp
Authorization: Bearer <api key>
streamable http, keys under settings then api keys. same tools.
the first three prompts worth running
prove the data is live
what accounts did tracked investors start following in the last 3 days, and which had more than one investor arrive?
if that returns accounts you have not heard of, the connection is doing its job. that is the convergence signal, and it is the thing the agent cannot get from a web search.
prove it can reason over it
take the top five and pull the founders for each. tell me which one has the strongest technical background for building in robotics, and show me the evidence you used.
this uses founders_batch rather than five separate calls, which is what keeps the run inside a sane time budget.
prove it closes the loop
for that founder, is there a warm path from anyone in my network? if yes, draft a short first message. do not mention how I found them.
that last instruction matters. an agent will otherwise write "I noticed several investors recently started following you", which is a good way to make a founder feel surveilled.
what to do when an answer looks wrong
the agent will occasionally state something with confidence that the tools did not return. the fix is to make it show its work:
which tool call did that number come from? show me the raw result.
typed tools make this checkable in a way that a web-search answer never is. if it cannot point at a call, it made it up.
related
open source deal flow
companies and founders surfaced on frontrun. curated, every week.
