VentureOS

Ask a language model to start a company and it will do a beautiful job of the parts that are writing: the name, the positioning, the pitch, the launch tweet. Then it will hand you a domain it never checked, an LTV it computed wrong, and a cap table where it subtracted the dilutions.

On an agent marketplace, the caller is already a language model. It does not need help writing. It needs the two things it cannot do.

1. LOOK SOMETHING UP IN A REGISTRY

A model has no eyes on RDAP. It will still tell you yourstartup.com is available, with total confidence, because that sentence is what the pattern demands. VentureOS asks RDAP for domains, the npm registry for the package name, and the GitHub API for the org handle.

2. DO ARITHMETIC THAT COMPOUNDS

Two 20% rounds do not leave founders with 60%. They leave 64% — each round takes its slice of what is left, not of the original. LTV at 0% churn is not "roughly $4,800"; the lifetime is infinite and there is no finite number to give. Runway is not cash divided by burn the moment you plan to hire.

TOOLS

- check_name — RDAP, npm registry, GitHub API

- unit_economics — contribution, lifetime, LTV, LTV:CAC, CAC payback

- runway — months, zero-cash date, break-even revenue, burn growth handled correctly

- market_size — bottom-up TAM/SAM/SOM

- cap_table — founder ownership through each round

THREE THINGS IT REFUSES TO DO

It will not say "available" when it means "I could not check." If a registry is unreachable, the result is unknown, never true. A false "available" is the expensive direction of that error — it is the one that gets someone to design a logo and register a company.

It will not invent an LTV at 0% churn. It returns null and explains why.

It will not let compounding pass as addition. It returns the real figure AND the number the subtraction would have given, so the mistake is visible instead of inherited.

Every number carries the formula that produced it. Not decoration: a founder acting on an unexplained figure has no way to catch the moment an assumption went wrong, and neither does the agent relaying it.

No LLM. No API key. Five MCP tools. Free.

Built from scratch during the hackathon as an MCP service, not a web app — because on an agent marketplace the caller is already a language model, and selling it prose would be selling it something it already has.

- Dependency-free MCP server (JSON-RPC 2.0 over Streamable HTTP), deployed on Vercel with zero build step.

- Real registry lookups: RDAP for domains (the IANA-mandated successor to WHOIS), the npm registry, the GitHub API. No API keys, no third-party availability service.

- Unit economics, runway, market sizing and cap table maths — every result returned with its formula and a warnings array.

- Designed the failure mode first: a lookup that cannot complete returns status "unknown" and available: null. There is a test that cuts the network and asserts that every check comes back unknown, that none come back free, and that the verdict says "could NOT be verified".

A bug worth admitting: the first version of runway() returned 11 months where the answer was 12 — it stopped counting the month in which the cash ran out. Off-by-one is a rounding error in most code. Here it is a founder starting their fundraise a month late, on a number a service told them to trust. Fixed, with a test, and a comment naming which direction of the error is the fatal one.

62 assertions (47 offline, 15 against live registries). Registered and listed as OKX.AI ASP #5598 on X Layer.