AgentScout Skill
AgentScout turns the live web into agent-ready input: give it a URL and it returns clean
markdown (read), structured JSON validated against your schema (extract), a free price
probe (quote), or a whole-site crawl (crawl). Every paid fetch settles in real USDC on
Base via the x402 protocol — no accounts, no API keys, no hand-rolled PAYMENT-SIGNATURE
signing. The service lives at https://api.agentx402.ai/v1/scout/*.
You hold one AgentScout identity, in one of two shapes (auto-detected by the client):
- Wallet-as-payer (the default): a signable EVM wallet pays each fetch inline via x402. AgentScout mints and manages a local wallet on first use — fund it and go, no sign-up.
- Account-key: for a managed wallet that can't sign, an opaque
ak_…bearer token identifies the account and debits prepaid credits. Credits are funded out-of-band via AgentKV (see Account-key mode & funding below).
Content is NOT encrypted. Unlike AgentKV — which encrypts values client-side so the server is zero-knowledge — AgentScout returns fetched web content in the clear. It is public web data by definition; there is no encryption key and nothing zero-knowledge about a scout response. Do not treat a scout result as private.
When to use AgentScout
Use AgentScout when you need:
- Grounding in live web content — pull a page as clean markdown (
read) to summarize, quote, or reason over, without a headless browser or an HTML parser of your own. - Structured extraction — turn a page into JSON that validates against a JSON Schema you
supply (
extract), e.g. a product's price/spec table or an article's title/author/date. - Free price discovery — call
quotefirst to learn the exact toll for a URL before spending anything;quotenever signs and never spends. - Site crawls — walk a site up to
max_pages(crawl) and get back a manifest of results plus artifact links; long crawls resume byjobId.
Do NOT use AgentScout for:
- Secrets or credentials. Scout fetches public web content; it is not a secret store (use AgentKV's secret-safe tools for that).
- Private pages behind auth. AgentScout fetches as an anonymous client — it cannot log in, carry your session cookies, or reach anything gated behind a login wall.
- Zero-budget contexts.
read,extract, andcrawleach cost real USDC; callquoteand check your spend cap before a loop or a bulk run.
Available MCP Tools
The agentscout MCP server exposes five tools — one free probe, three paid fetch verbs, and
a free crawl-status check:
| Tool | Description | Cost |
|---|---|---|
scout_quote | Free toll-price probe for a URL — returns atomic-USDC prices and a would_pay advisory. Never signs, never spends. | Free |
scout_read | Fetch a URL and return clean markdown. fresh bypasses the ~6h read cache; max_tokens truncates the result. | ~$0.003 USDC + publisher toll |
scout_extract | Fetch a URL and extract structured JSON validated against a JSON Schema you pass, with optional natural-language instructions. | ~$0.012 USDC + publisher toll |
scout_crawl | Crawl a site up to max_pages (REQUIRED, price-determining) and return a results manifest with artifact links. Long crawls return a jobId. | ~$0.002 USDC × max_pages + tolls |
scout_crawl_status | Check a crawl job's status by jobId — returns the manifest when complete, or the in-flight status. | Free |
Publisher tolls are wallet-mode only. A page may carry a publisher toll on top of the base fee — extra USDC the fetch fronts to the site.
max_toll_usdcaps that toll, but it only applies in wallet mode: anak_(account-key) caller cannot front a toll, so settingmax_toll_usdin account-key mode is refused client-side (tolls_require_x402) before any request is sent.
Paying & spend caps
Three ceilings bound what a scout session can spend. Set them in the plugin config (run
/plugin to edit) or as env vars for the CLI:
maxSpendUsd(AGENTSCOUT_MAX_SPEND_USD) — refuse any single fetch that would cost more than this.maxSessionSpendUsd(AGENTSCOUT_MAX_SESSION_SPEND_USD) — refuse further fetches once cumulative spend across the whole session exceeds this.maxTollUsd(AGENTSCOUT_MAX_TOLL_USD) — cap the publisher toll on a paid fetch. Wallet mode only (see the caveat above).
Caps refuse — a paid tool over the ceiling throws a SpendCapError rather than silently
capping and spending less. The SDK always signs the server's exact 402 challenge amount, never
a self-computed one, and enforces every cap before signing. scout_quote is free and never
counts against a cap, so probe first, then spend within budget.
Account-key mode & funding
Account-key mode uses an ak_… bearer token instead of a signing wallet and debits prepaid
credits. Two things are specific to scout:
-
Scout has no deposit route. AgentScout does not mint or fund accounts — there is no
scout/depositendpoint.ak_credits are funded out-of-band via AgentKV, which shares the same account ledger:# Fund the ak_ account from ANY signing wallet (via AgentKV's deposit route): awal x402 pay https://api.agentx402.ai/v1/account/deposit \ --headers '{"Authorization":"Bearer ak_..."}'The credits that deposit buys are then spendable by AgentScout under the same
ak_bearer. -
Tolls are refused in account-key mode. A managed account cannot front a publisher toll, so setting
max_toll_usd(maxTollUsd) with anak_credential fails fast client-side with atolls_require_x402error — before any request is issued. Use wallet mode if you need to pay tolls.
Crawl: resumable jobs
A crawl can outlast a single request. If it does not finish within timeout_ms, scout_crawl
returns a pending handle carrying a jobId instead of blocking:
scout_crawl { url, max_pages, timeout_ms }→ either a completed manifest or{ jobId, … }.- Poll
scout_crawl_status { job_id }(free) until it returns the completed manifest. - Fetch individual page artifacts referenced in the manifest via the SDK/CLI
(
agentscout crawl artifact <jobId> <key>).
Because max_pages is price-determining, the crawl settles its full cost upfront
(max_pages × per-page price); resuming a job by jobId is free.
One-Time Setup
Install the plugin from Claude Code's marketplace — the exact /plugin marketplace add and
/plugin install commands are in the plugin's README.md (plugin/README.md). After install,
Claude Code prompts for the config and threads it into the MCP server for you.
1. Credentials (entered at install, not via shell env)
| Config | Required | Description |
|---|---|---|
| Wallet private key | No | Optional — leave blank and AgentScout mints + manages a local wallet on first use (then fund it). To bring your own: an EVM private key (hex), the wallet that pays. |
| Account key | No | An ak_… bearer token for managed-wallet (credit) mode. Funded out-of-band via AgentKV. Mutually exclusive with a wallet key. |
| AgentScout endpoint | No | The hosted API; defaults to https://api.agentx402.ai. |
| Network | No | eip155:8453 (Base mainnet, default) or eip155:84532 (Base Sepolia testnet). |
| Max per-operation spend (USD) | No | Refuse any single fetch that would cost more than this; empty = no per-op cap. |
| Max session spend (USD) | No | Refuse fetches once cumulative session spend exceeds this; empty = no cap. |
| Max publisher toll (USD) | No | Cap the toll on a paid fetch. Wallet mode only; rejected in account-key mode. |
Re-run /plugin to change these later. Verify the server loaded with /mcp (you should see
the agentscout server and its five tools).
2. Fund the payer
- Wallet mode: send USDC to your wallet address on Base — reads/extracts/crawls are then paid inline per fetch via x402. Don't have a wallet? Leave the private-key config blank — AgentScout mints and manages a local wallet on first use — then fund the address it prints.
- Account-key mode: deposit to AgentKV's
/v1/account/depositunder yourak_bearer (see Account-key mode & funding above) — scout debits those shared credits.