We build tools agents can trust.
Autonomous software needs to call something, get an answer, and act on it — with no human reading the output first. That only works if the tool behaves identically every time, costs a knowable amount, and says what it could not determine.
Three properties, chosen for how an agent fails.
Deterministic
No model in the response path. The same transaction always produces the same JSON — nothing to hallucinate, nothing to drift.
No accounts, no keys
Paid per call in USDC over x402. An agent can discover, pay and retry entirely on its own.
Honest about gaps
Every response says which checks ran and which could not. An empty result is never mistaken for a clean one.
One product, and the two fields that make it different.
base-transaction-decoder explains any Base mainnet transaction as strict JSON — plain English, a typed action, every asset moved with amounts already scaled, counterparties labelled, gas in dollars, risk flags with evidence.
provenance closes.$0.02 per call in USDC · 50 free calls per IP address per day · no account, no signup, no API key
We deployed a token, then took it apart in public.
On 2026-09-01 we deployed a token on Base. Within a day we examined its structure and retired it — the pool withdrawn, the entire supply sent to an address it cannot leave. Two holders at the time: our own wallet and the pool. Nobody outside held it.
The reasons, the arithmetic, both transaction hashes and a verification anyone can run are published permanently. The record of what was wrong with it was not removed — including two dated corrections to things we got wrong, with the mistaken wording still visible above each fix.
Practices, not principles.
A practice can be checked. A principle cannot.
- Claims carry the date they were verified. A fact true last week is reported as a fact about last week.
- Pages are never overwritten. Corrections are dated and the original stays visible.
- Examples are executed before publication. Every command shown here is run against production.
- Unreachable is not clean. A source that cannot be checked is reported as unknown, never as an absence of problems.
- What could not be determined is stated alongside what was. The same rule the decoder follows in its own output.
One person, and gates that fail loudly.
0200project is operated by one person, with AI systems doing the majority of the engineering, review and analysis. Every claim is checked against a source before it ships, and the checks are automated wherever they can be.
Read from production, right now.
These numbers are fetched from our own health endpoint when this page loads. They are not written into the page and we cannot edit them without changing what the server reports.
If this panel fails to load it will say so rather than showing nothing — an empty panel and a broken one look identical, and only one of them is honest.
Talk to us.
General enquiries and corrections: contact@0200project.com. Bugs and feature requests go to GitHub Issues. Security reports have their own page.