Documentation

Connect your agent

Two ways in. Use MCP if your agent already speaks it — it gets discovery and payment in one place. Use plain HTTP if you want to see every byte. They run the same code and charge the same amount.

Everything below is on Base Sepolia with test USDC. Your agent needs a wallet holding test USDC and nothing else — the gas is paid by the facilitator, not by you.

Option A

Over MCP

Add the server to your agent's MCP configuration. In Claude Code that is .mcp.json; other clients use the same shape.

{
  "mcpServers": {
    "aiskills402": {
      "url": "https://mcp.aiskills402.com/mcp"
    }
  }
}

Two free tools appear, plus one paid tool per skill:

search_skills — the catalog, with the price of each skill.
get_skill — one skill: price, input limits, output shape.
execute_humanize_long — runs Humanize Long. Costs $0.08 and needs a wallet.
execute_humanize — runs Humanize. Costs $0.03 and needs a wallet.

Each skill gets its own paid tool rather than one shared execute, so your agent sees the price in the tool list before calling anything.

Your agent must be able to sign x402 payments. Without a wallet the paid tool answers with a 402 payload describing what it wants; with one, the client signs and repeats automatically.

Option B

Over plain HTTP

Discovery is free and needs no key:

curl https://api.aiskills402.com/v1/skills
curl https://api.aiskills402.com/v1/skills/humanize-long

Calling the endpoint without paying returns 402 Payment Required and, in accepts, everything needed to pay: amount, asset, network and recipient.

curl -X POST https://api.aiskills402.com/v2/skills/humanize-long/execute \
  -H "Content-Type: application/json" \
  -d '{"input": "<your text>"}'

Sign one of the offered requirements, base64 the payload, and repeat the request with it in X-PAYMENT. The response is 200 with the result and the transaction hash.

curl -X POST https://api.aiskills402.com/v2/skills/humanize-long/execute \
  -H "Content-Type: application/json" \
  -H "X-PAYMENT: <base64 signed authorization>" \
  -d '{"input": "<your text>"}'

Any x402 client library does the signing for you — the reference client is documented at x402.org.

The money

What you are charged, and when

The amount is fixed and shown before anything runs. It is taken only after the work succeeds — we verify the signature, do the work, and then settle. If the run fails, the authorization is cancelled and no transfer happens.

Skills are priced by how much text they accept. The cost of a rewrite grows with the length, so a single price for every size would either overcharge a short email or sell a long article at a loss. Pick the smallest one that fits your text.

Skill	Takes	Price
humanize-long	up to ~5,100 words	$0.08
humanize	up to ~1,300 words	$0.03

Outcome	You are charged
Success	the skill's price
Run failed	nothing
Cost cap hit mid-run	nothing
Input rejected before the run	nothing — 400 before the 402
Skill paused or unavailable	nothing — 503

Every response carries settled. Treat that field, not the HTTP status, as the answer to “did I pay”.

When it says no

Errors your agent will meet

Code	What to do
input_too_short	Send at least 300 characters. Shorter text is not worth rewriting.
input_too_large	Use a skill that takes longer text, or split it. The response tells you the limit.
input_required	The body needs an `input` field holding a string.
skill_not_found	Check the slug against /v1/skills.
skill_unavailable	Paused or between versions. Retry later — you were not charged.
payment_replay	That authorization was already used. Sign a fresh one.
budget_exceeded	The run cost more than its cap and was stopped. Not charged.
translated	The model answered in the wrong language, so the run is void. Not charged.

Your text

What we keep

By default, nothing readable. We store a SHA-256 fingerprint of the input and of the output — enough to prove later which input produced which output, without us holding your text.

Send "retain": true and a full copy is kept for 30 days and then deleted automatically. Failed runs are never kept, even with the flag set.

Honestly

What this does not do

The rewrite is judged on reading like a person wrote it, with every factual claim preserved and the length kept. It is not sold as a way to get past AI detectors, and we measured why: the same guarantees that keep your facts intact are what detectors recognise. Anyone promising both is promising one of them falsely.

Payments are on a test network. Nothing here moves real money yet.