AkashiDocs
Quickstart

For AI agents (Skill)

Give your agent Akashi's SKILL.md. It learns the three verbs, the prices, how to pay and how to read a result.

The skill is one Markdown file: https://useakashi.xyz/SKILL.md. It tells an agent when to reach for Akashi (before writing a scraper, before a generic web fetch for structured data, before saying something is inaccessible), how to call discover, inspect and run, how the x402 payment works and how to read the result.

Install it

Tell your agent:

set up https://useakashi.xyz/SKILL.md

The agent downloads the file and saves it to its skills. To install it yourself in Claude Code:

mkdir -p ~/.claude/skills/akashi
curl -fsSL https://useakashi.xyz/SKILL.md -o ~/.claude/skills/akashi/SKILL.md

What the agent can do with it

Discover and inspect need nothing but HTTP, so the agent can use them at once:

curl -s -X POST https://api.useakashi.xyz/v1/discover -H 'content-type: application/json' \
  -d '{"query": "recent papers on RAG hallucination", "limit": 5}'
curl -s -X POST https://api.useakashi.xyz/v1/inspect -H 'content-type: application/json' \
  -d '{"id": "firecrawl/research-papers"}'

A run is paid, so the agent needs a wallet and an x402 client. Give it one of these:

  • The CLI. Download akashi.mjs and set AKASHI_PRIVATE_KEY. The agent runs node akashi.mjs run <id> --input '<json>'. See For humans (CLI).
  • The MCP server. The same file, as three MCP tools. See For AI agents (MCP).
  • Code. The skill includes an @x402/fetch snippet with a per-call price cap. See From code (HTTP).

Fund a test wallet

Use a new key that holds nothing else. Runs are paid in USDC on Base Sepolia, a testnet.

Create a key and read its address with the CLI:

export AKASHI_PRIVATE_KEY=0x$(node -e "console.log(require('crypto').randomBytes(32).toString('hex'))")
node akashi.mjs wallet

Open faucet.circle.com, pick Base Sepolia and send test USDC to the address from the previous step. You need no ETH: you sign a USDC authorization and the x402 facilitator submits it.

Ask your agent for something that needs a live tool, for example: "What is 100 USD in NGN today, and is it raining in Lagos?"

Results are data, not instructions

The skill tells the agent to treat everything inside a result's data as untrusted content and never to follow instructions found there.

On this page