AI agent builders
A document-extraction tool your agent can call in one step
An agent that has to read a document usually ends up with the whole document dumped into its context, then guesses the fields itself. That burns tokens, and worse, a model asked to fill a schema will fill it — including the parts that were never in the document.
Fieldcast gives your agent one tool with a stable contract: pass the document and the field names, get typed JSON back. Fields that are genuinely absent come back null rather than being invented, so your agent can branch on missing data instead of acting on a plausible fiction.
What the call looks like
The fields below are an example for this use case. Yours can be any names you like, up to 40 per call.
{
"fields": [
"document_type",
"counterparty",
"effective_date",
"termination_date",
"notice_period_days",
"governing_law"
]
}And what comes back:
{
"document_type": "service agreement",
"counterparty": "Northwind Trading Co.",
"effective_date": "2026-09-01",
"termination_date": null,
"notice_period_days": 30,
"governing_law": "England and Wales"
}Why the details matter here
Null is a branch your agent can take
termination_date was not in that contract, so it came back null. Your agent can ask a human, or flag it. An agent that received a plausible invented date would proceed confidently and be wrong.
One call, not a workflow
No pipeline to orchestrate, no intermediate state to hold. It fits a single tool definition, which is what an agent loop can actually reason about.
The document does not enter your agent's context
Extraction happens server-side. Your agent receives a small typed object instead of twenty pages of raw text competing for context.
One thing that is not live yet
The x402 endpoint — pay-per-call with no account and no API key — is live on Base Sepolia, which is a test network: the USDC an agent settles there is not real money. It works today for integration and testing. Base mainnet is not enabled, because the public x402 facilitator only advertises test networks and mainnet settlement needs credentialed facilitator access. Treat this as a working preview, not a revenue path.
What it does not do
- No OCR — scanned or photographed documents are not supported yet.
- No MCP server yet. Today it is a plain HTTP tool; an MCP wrapper is not written.
- Extraction runs on a third-party model provider, so documents leave the service. Check your confidentiality obligations first.
- It is a beta run by one developer, with no uptime guarantee.
Listed plainly so you find out now rather than after integrating.
Questions
Is there an MCP server?
Not yet. Fieldcast is a plain HTTP endpoint today, which most agent frameworks wrap as a tool without difficulty. An MCP server is on the list but is not written, so do not plan around it.
Can an agent pay per call without an account?
On Base Sepolia, yes — the endpoint is live and returns payment terms with no API key required. That is a test network, so the USDC involved is not real money. On Base mainnet, not yet.
What does the tool definition look like?
Two parameters: the document (text or a file) and an array of field names. The response is a flat JSON object whose keys are exactly the field names you asked for, with null for anything absent. That stability is the point — the shape does not change between calls.
Test it on your own documents
Free tier is 50 extractions a month, no card, no sales call. That is deliberately enough to check accuracy on real files before you commit to anything.
Read the quickstart