Curated data sources

Making a call

Every method is forwarded. /v1/data/:source/* passes the HTTP verb, the query string (including repeated parameters, which arrive upstream as a repeated key) and the JSON body straight through to the operation the path resolves to.

Authenticate with Authorization: Bearer <key>. Mint a key in the dashboard at app.superagnt.com.

A request body is JSON: send it with Content-Type: application/json, because the content type you send is passed upstream with the body.

curl "https://api.superagnt.com/v1/data/<source>/<endpoint>?limit=25" \
  -H "Authorization: Bearer $AGNTDATA_API_KEY"

const res = await fetch(
  "https://api.superagnt.com/v1/data/linkedin/get-company-details?username=microsoft",
  { headers: { Authorization: `Bearer ${process.env.AGNTDATA_API_KEY}` } },
);
const body = await res.json();

if (!body.success) {
  throw new Error(`${body.error.code}: ${body.error.message}`);
}

console.log(body.data);           // LinkedIn's payload, in LinkedIn's own shape
console.log(body.meta.costCents); // what this call was charged, in cents

What you pay

Every successful response carries its own price. meta.costCents is the amount actually charged for that call, after any discount on your account, and the two remaining balances sit beside it. You never need a second round trip to find out what a call cost or what is left.

{
  "success": true,
  "data": { "...": "the upstream payload" },
  "meta": {
    "costCents": 1,
    "purchasedBalanceCents": 4820,
    "subscriptionRemainingCents": 1500,
    "cached": false,
    "latencyMs": 412
  }
}

data is the source's own payload, in that source's field names: there is no common schema across sources. For a compact view of it instead, a markdown table or only the fields you name, see Response views and paging. purchasedBalanceCents and subscriptionRemainingCents are what your data wallet has left after the call: the purchased balance and the remaining subscription allowance, both explained in Credits and wallets. latencyMs is how long the upstream call took, retries included, and cached is always false.

Pricing is set per slug, with optional per-endpoint overrides where one operation is materially more expensive than the rest of its source. Some sources are variable-cost by nature: the web and seo sources price a call by what the request actually did, so the platform bills the real upstream-reported cost rather than the pre-call estimate. The estimate is only the balance gate.

Failed calls and empty results cost nothing. An upstream error, a timeout, and a failure reported inside an HTTP 200 are billed at zero and logged as errors. A query that matched nothing is a success at zero cost, with meta.empty set.

Pagination

Curated sources keep each upstream's own pagination convention. Most commonly that is a dedicated continuation operation: you call the list endpoint, take the continuation token out of the response, and feed it to the matching …_continuation operation. Others page by cursor, by number or by offset.

There is no single platform-wide cursor on this surface, so check the endpoint you are calling in the reference rather than assuming. Elsewhere on the platform, the two webhook delivery lists page by cursor, and the workspace database and /v1/usage by limit and offset.

Response views and paging walks through each of these paging styles over MCP and over REST.

From an agent

Over MCP the same operations appear as tools named data_<slug>_<operationId>, for example data_reddit_search_posts. Same validation, same pricing, same credit balance as REST.

On Claude, add the server as a connector: open the prefilled dialog on claude.ai and click Add. Claude Code signed in with an API key, or another CLI, connects with:

claude mcp add --scope user --transport http superagnt https://mcp.superagnt.com/mcp
claude mcp login superagnt

Curated sources are on by default on a server that lists grouped tools, as every workspace MCP server does unless it was set up per operation on purpose: a few grouped tools per source keep the tool list short. A per-operation server keeps them opt-in: find a source with agnt_tools_search, then turn it on with agnt_tools_enable and the family id data:<slug>, for example { "families": ["data:x"] }. The first-party agnt source is on every server and never needs an enable.

Migrating off the pre-/v1/data path