About

Three kinds of resources

Agents need to find different things: A2A agents tell an agent who can do a job, x402 lets it pay per call, and MCP servers give it the tools to do the work. We index all three behind one search and one set of MCP tools, browsable at /services, /mcp and /a2a.

What is x402?

x402 is Coinbase's HTTP 402 micropayment protocol. AI agents can discover a service, inspect its payment requirements, and then call the external API with an x402 payment flow. We index public x402 endpoints; agent-tools.cloud itself is discovery-only.

Agent-friendly discovery

Agents can use POST /api/v1/ask for intent-level recommendations, GET /api/v1/search for faceted search, and GET /api/v1/services/{slug} for a full service card.

Discovery entry-points: /.well-known/agent-tools.json · /.well-known/mcp.json

Safety scanning

Every indexed MCP server is re-scanned each hour for known malware and abuse patterns — pipe-to-shell install lures, base64 / eval download cradles, bare-IP command-and-control URLs and prompt-injection phrasing in its advertised tool metadata. It also detects MCP tool-poisoning coercion — descriptions that try to hijack an agent's tool-calling (“always call this tool first”, “before using any other tool you must…”), hidden <IMPORTANT> instructions, demands to list API keys or include secrets in a response, and coercion to read & forward .key/.pem/.ssh/.env files. Each server carries a clean / suspicious / malicious verdict with the matched rules. Scanning is static only — we never execute server code.

You can also scan any MCP endpoint on demand, before you connect: call the scan_mcp_safety tool on the MCP server at /mcp-discovery, or send its URL to our A2A agent at /a2a. If the server is already indexed you get our latest stored verdict instantly; if not, we probe it live, scan it, and add it to the directory. Each on-demand call also runs a frontier large model second-opinion as an advisory dimension alongside the deterministic rules.

Grading

Every listing carries a 0–100 score and a letter grade, built only from signals we measure ourselves. We do not re-export another directory's rating.

x402 services

Availability	40	Endpoint probe: ok 100%, degraded 45%, unknown 22%, down 0.
Payability	20	Real HTTP 402 challenge 15, /.well-known/x402 descriptor 5, payment address 5, settlement network 3, price 2 — of 30, scaled to 20.
Demand	40	30-day USDC receipts at the payTo address, read from the chain. Distinct payers carry two thirds, transfer count one third; both on a log scale. Chains with no free indexer (currently Sui) leave this blank rather than zero.

A service whose endpoint is down is held at D regardless of the other two components.

MCP servers

Availability	25	ok 25, degraded 10.
Conformance	25	JSON-RPC initialize + tools/list: pass 25, partial 10.
Performance	30	Full marks at p95 ≤ 50 ms, falling linearly to 0 at ≥ 2000 ms. No latency data scores 0.
Descriptor	20	Advertises its tools 8, tool metadata captured 5, source or package published 4, transport declared 3.

A server flagged by our safety scan is held down regardless of the four components above: suspicious caps at C, malicious at E. The matched rules are shown on the server's page.

A2A agents

Availability	35	Agent card fetched and parsed: ok 35, degraded 14.
Conformance	25	pass 25, partial 10.
Card detail	20	Skills declared 8, three or more skills 4, callable endpoint 6, capabilities 4, protocol version 3, auth scheme 3, documentation 2 — of 30, scaled to 20.
Performance	20	Full marks at ≤ 150 ms, falling linearly to 0 at ≥ 1500 ms.

Grade bands

Bands are set per resource type, so a grade means "top of its own kind" — the three scorers measure different things and are not directly comparable.

Type	A	B+	B	C+	C	D
x402	88	80	62	53	43	12
MCP	95	88	72	56	40	20
A2A	98	96	92	85	66	60

Anything below the D threshold is E.

Re-scanning

Availability is re-probed every 4 hours, on-chain demand daily, and payment metadata on each verification pass. Grades recompute automatically after every pass — there is nothing to request. Each x402 listing page shows its own breakdown with the timestamps of the last two checks, and says why a component reads zero. The most common cause is that we could not extract a payment address from your 402 response or descriptor; publishing accepts[] with payTo, network and an amount fixes both that and the payability component.

Claiming a listing you operate

Almost everything here was crawled, so most listings describe someone who never asked to be indexed. If one of them is yours, you can take it over and correct it: sign in with GitHub, then publish a token we issue on the host that serves the listing — in a descriptor you already publish, in a well-known file, or as a DNS TXT record.

Nothing is required of you. Listings stay up either way and submitting one does not need an account. A verified claim gets an Owner verified badge, lets you edit the descriptive fields, and adds 3 points to the grade. Measured fields — availability, grade, latency, conformance, on-chain demand — are not editable by anyone.

How to claim a listing →

Sources

• · MCP servers — Smithery, the official MCP registry, PulseMCP, Glama
• · x402 — awesome-x402, x402scan
• · Manual submissions (/submit)

Retired paid relay

The previously hosted Qwen relay and crypto vertical endpoints were retired on 2026-05-25. All paid relay routes now return 404; this host now serves the directory and MCP discovery layer only.