Hosaka

deadchannel & Hosaka

Two services that sell to AI agents over x402, paid per call in USDC. No signup, no API key, no subscription — an agent calls, pays, and gets an answer.

Agents can install Hosaka as a skill, which tells them what it is for and when to reach for it:

npx skills add plus8bit/deadchannel/skills/hosaka --yes

Hosaka — company data for agents

Available as an MCP server, so any MCP client can buy from it directly:

npx -y hosaka-mcp

{  "mcpServers": {    "hosaka": {      "command": "npx",      "args": ["-y", "hosaka-mcp"],      "env": { "HOSAKA_PRIVATE_KEY": "0x…" }    }  }}

HOSAKA_PRIVATE_KEY is a wallet holding a little USDC on Base. Four tools at four prices, so a cheap question never pays for an expensive answer:

Why the dossier is worth having. A company proves it owns its domain to every SaaS product it buys by placing a DNS verification record, and authorises every sender it uses in its SPF record. Those two lists are a purchase history the company published itself.

Asking for figma.com returns Anthropic, OpenAI, Adobe, Atlassian, MongoDB Atlas, Greenhouse, Docusign, Stripe, Notion, Dropbox and Zendesk — each with the exact record that proves it, so a buyer can check rather than trust.

Page fingerprints require a loaded script or CDN host, never a mention: a site listing a vendor's logo among its integrations is not a site that uses it.

Live at hosaka-agents.vercel.app.

deadchannel — a risk oracle for x402 endpoints

Give it a URL, it tells an agent whether that endpoint is alive, honestly priced, and safe to call — without spending a cent.

Roughly 17% of probed x402 endpoints are dead or traps, and the Bazaar discovery layer ships no ranking, quality, or reputation signal by its own admission. An agent picking from that catalog is guessing with real money.

What it checks

Every verdict comes from unpaid 402 responses, which is what makes it cheap enough to run across a whole catalog.

Verdicts: live, degraded, trap, testnet, dead, unknown, plus a bounded 0–100 risk score.

Use

Requires Node 22+. There are no runtime dependencies — types are stripped natively, so there is no build step and nothing to audit but this repo.

node src/cli.ts https://x402.org/protectednode src/cli.ts --quiet --samples 5 url-a url-b url-cnode src/cli.ts --json url > verdict.json

Exit code is non-zero when any target is a trap or dead, so it drops into CI as a guard against shipping an agent pointed at a bad endpoint.

TESTNET   risk 80   $0.01  355ms p99  https://x402.org/protected  x network-mainnet   Only testnet networks offered (base-sepolia, solana-devnet).                      This endpoint cannot accept real value.  ! schema-advertised No input or output schema. An agent has to pay before it                      can find out what it gets back.

First full catalog scan — 20 Aug 2026

node src/scan.ts --live 150 pulls every resource the public Bazaar facilitators publish and audits it. Results over 14,979 resources:

Median price $0.01, range $0 to $1,000, 316,927 paid calls in 30 days.

The catalog is in better shape than the folklore suggests. What it is not is evenly distributed: three addresses list 2,750 resources between them and receive one call in eighty, which makes them roughly 15x over-represented relative to the demand they serve. Catalog size is not catalog depth.

[catalog audit]

A correction, and the guard against repeating it

An earlier version of this table claimed 89.8% of the catalog published no tags. That was wrong. The loader read extensions.bazaar.tags only, while most publishers put tags on the item root — the real figure is 40.9%.

The bug survived review because nothing asserted the positive count. A check that only ever counts what is missing cannot tell "publishers omit this" apart from "we are looking in the wrong place." test/catalog.test.ts now hand-counts tags across all three known locations in a captured 200-item slice of the live catalog and asserts the loader matches exactly, in both directions.

Notes from the wild

Built against live endpoints, not the spec alone:

• v2 moves payment requirements into a base64 payment-required header and leaves the body as {}. A parser that only reads the body marks the official reference endpoint as dead.
• v2 sends CAIP-2 network ids (eip155:84532, solana:EtWTRA…) where v1 sent friendly names. Both are live simultaneously.
• v2 renames maxAmountRequired to amount and hoists shared resource metadata to the payload root.
• Several servers omit the top-level x402Version the reference implementations all send, and some nest accepts[] one level deeper than documented. Both are parsed and reported as warnings rather than rejected.
• Most resources are POST. Probing them with GET returns 404/405, which looks exactly like a dead endpoint — it put our first live-probe dead rate at 25% when the real figure is 2.7%. The verb comes from the catalog now, with a POST retry when it is unknown. Any x402 index reporting a high dead rate is worth checking for this.
• Discovery metadata is scattered. The spec puts serviceName/tags on the ResourceInfo object, the CDP catalog flattens them onto the item root, and a minority nest them under extensions.bazaar. Read one location and you will undercount by half.
• Brokered rails exist. AWS Marketplace resources name the payee with a URN under an aws:base network instead of a chain address. That is legitimate, but the funds go to the broker, so it is reported as a warning rather than scored as an invalid payout.

Selling

The service sells its own check over x402. src/server/ implements the seller side of the v2 HTTP transport: PAYMENT-REQUIRED out, PAYMENT-SIGNATURE in, PAYMENT-RESPONSE back, in the authorization flow — verify, run the resource, then settle.

No private key is involved. A seller declares where settlement should land; the buyer signs and the facilitator broadcasts. The payout address is therefore public information and lives in deadchannel.config.json, where anyone can audit it, rather than in a dashboard where nobody can.

Settlement runs only after the probe produced a result, so a failure on our side costs the buyer nothing.

GET  /             service card, freeGET  /health       liveness, freeGET  /facilitator  proves credentials are accepted, free, moves no moneyPOST /probe        the check, $0.005 in USDC on Base

GET /facilitator exists because a wrong credential otherwise stays invisible until someone tries to pay, and the first to discover it would be a customer.

Live

The service is deployed at https://deadchannel.vercel.app, selling on Base mainnet at $0.005 per call and settling through the Coinbase facilitator.

First settled payment: 0x6ac4a22c, block 50230005, 20 Aug 2026. Gas was paid by the facilitator, not the buyer — in x402 the buyer only signs, so a wallet holding nothing but USDC can pay.

npm run validate re-runs the 25 preflight checks the Bazaar applies before it will index a resource, and reports whether the listing is currently active. It needs no key and moves no money.

resource : https://deadchannel.vercel.app/probevalid    : trueaccepted : acceptedchecks   : 25/25 passedindexed  : active=true

Hosaka

A second shop on the same payout address, selling company facts to agents.

POST /lookup    $0.005  domain age, registrar, mail and DNS provider, DMARC, HTTPSPOST /dossier   $0.02   every third-party vendor we can prove, with the proof

Live at https://hosaka-agents.vercel.app, settling on Base mainnet.

Profiles are assembled from four sources that need no key and no supplier: DNS over HTTPS, the RDAP registry, the TLS handshake and the homepage. Cost per profile is zero — the margin is in the assembly.

The interesting part is vendor detection. A company proves ownership to every SaaS product it buys by placing a DNS verification record, and authorises every sender it uses in its SPF record. Those two lists are a purchase history the company published itself. Every claim carries the record that proves it, so a buyer can check rather than trust.

Using it from Claude or ChatGPT

The Bazaar's search ignores the query and ranks by unique payers, so a new shop is invisible there regardless of quality. MCP is the channel that works on day one — the agent is already in a client that speaks it.

{  "mcpServers": {    "hosaka": {      "command": "node",      "args": ["/absolute/path/to/src/hosaka/mcp/server.ts"],      "env": { "HOSAKA_PRIVATE_KEY": "0x…" }    }  }}

The buyer brings their own wallet: the key signs the payment locally and never leaves the machine. Without it the tools say what is missing instead of failing mid-conversation.

Development

npm install     # typescript only, for typecheckingnpm test        # 18 tests, no network requirednpm run typecheck

License

MIT