Three platforms are racing to answer the same question: when an autonomous agent needs to find, verify, and pay another agent for an API call — who handles the introduction? With Claude Code auto mode going default-on August 14, that question stops being theoretical.
In the past 60 days, three platforms have emerged to solve the same coordination problem for autonomous agents:
| Platform | Listing Count | Model | Trust Mechanism | Free Tier |
|---|---|---|---|---|
| minia2a | 306 services | Pay-per-call marketplace | Verify-first (health probes) | 15 free trials/endpoint |
| vibes-coded | 298 skills | Agent tool API | Artifact risk scoring | Unclear |
| Aidress | 50 agents | Open-source registry (MIT) | Trust score (0–100) | Open source, self-host |
All three are building what Y Combinator calls "software for agents" (RFS). Aidress has the YC backing. vibes-coded has the most GitHub-native approach. minia2a has the most real-world usage.
The HN discussion of Claude Code auto mode (275 points, 302 comments) is 100% about safety, sandboxing, and permission fatigue. Zero percent is about what agents will do with all that autonomy. That's not a flaw in the discussion — it's a gap in the infrastructure. Agents can't buy things yet because the coordination layer doesn't exist.
Here are 10 categories of pay-per-call APIs your auto-mode agent can use right now, each with a real endpoint and what it returns:
x402-search · 118 calls · 6 users
Web search with structured results. Agent researches a topic → calls search → gets ranked results with snippets.
x402-gas · 1,644 calls · 218 users
Multi-chain gas prices. Most-used endpoint. Your agent checks gas before submitting transactions.
x402-polymarket · 869 calls · 101 users
Prediction market odds. Agent evaluates a claim → checks what the market thinks.
x402-token-security · 154 calls · 43 users
Token contract audit. Agent encounters a token address → verifies it's not a honeypot.
x402-screenshot · 381 calls · 24 users
Page screenshots. Agent needs to see a rendered page → captures it headlessly.
x402-email · 107 calls · 25 users
Send emails programmatically. Agent completes a task → emails the result.
x402-sentiment · 174 calls · 33 users
Sentiment analysis. Agent reads text → classifies tone and emotion.
x402-recall · 1,991 calls · 120 users
Data enrichment & recall. Second most-used endpoint. Agent has partial data → fills in the gaps.
x402-uuid · 348 calls · 41 users
ID generation, encoding, validation. Agent needs a unique ID → generates one that won't collide.
x402-find · 1,292 calls · 100 users
Intelligence lookup. Agent needs to know something → queries structured knowledge base.
Call counts and user numbers from live /api/stats, August 11, 2026. Top 3 by usage: gas (1,644), recall (1,991), find (1,292).
Everyone is building payment rails:
Five payment rails. Zero discovery standards.
When your auto-mode Claude agent wants to call an API, it faces a chicken-and-egg problem:
This is exactly what HN user greenfish6 described: "My Claude Code and my Codex agent always stops before I'm actually able to buy." It's not a payment problem — it's a discovery and trust problem. The agent doesn't know what's on the other side of the 402 response.
How does an agent decide whether to pay an unknown counterparty?
Aidress assigns every agent a trust score (0–100) based on transaction outcomes. After each call, the consumer rates the provider 1–10. Anti-gaming rules prevent self-dealing. 50 registered agents, all "verified," average trust score 75. But with only 67 total transactions across the entire registry, the scores are based on almost no data. Trust without volume is a prior, not a measurement.
vibes-coded assesses the risk of the output artifact — what the API returns. This is a different axis: not "is this agent trustworthy?" but "is this response safe to act on?" 298 skills, but the scoring methodology is opaque.
minia2a sends automated health probes to every registered service. If an endpoint returns HTTP 000 (unreachable) for a sustained period, it's deactivated. If it returns 200 but the response doesn't match the declared schema, it's flagged. This is mechanical verification — no ratings, no reputation, just "does this thing actually work right now?" Combined with trial-first (15 free calls per endpoint, no wallet required), the agent can test before trusting.
Even with discovery solved, there's a final trust problem: the agent needs permission to spend money. This is what greenfish6's complaint is really about. Claude Code stops before buying because it has no budget framework.
The .agent-budget proposal is a simple JSON file the developer puts in their project root:
{
"daily_limit_usdc": 5,
"max_per_call_usdc": 1,
"allowed_chains": ["base"],
"allowed_tokens": ["USDC"]
}
The agent framework reads this file. When it hits a 402 response with machine-readable x-402-amount and x-402-chain headers (which minia2a supports), it checks the budget. If within limits → auto-pay. If not → ask the human. This turns "always stops before buying" into "buys when it makes sense, asks when it doesn't."
On August 14, Claude Code auto mode becomes the default for all Pro, Max, and Team users. Millions of developers will have agents that can execute commands, read files, and make HTTP requests without asking for each step.
What happens when one of those agents tries to call an external API?
Today: it can't. The API returns 402 Payment Required, and the agent has no way to understand what that means in machine terms. It stops.
With the right infrastructure:
/.well-known/x402 → discovers available endpoints?probe=1) → gets price, chain, and token info in headers.agent-budget → $0.05 is within the $1/call limitNo human in the loop. No permission dialog. No "always stops before buying."
This is not science fiction. The pieces exist today. They just need to be connected.
15 free calls per endpoint. No wallet. No registration. The agent can verify the API works before committing money. Neither Aidress nor vibes-coded offer this.
Automated health probes deactivate dead endpoints. In a landscape where independent audits found 76% of x402 endpoints dead, this is table stakes.
Works with x402, Cloudflare Wallets, or PayAI. Not locked to one payment rail. If a new facilitator emerges tomorrow, it works.
Every 402 response includes x-402-amount, x-402-chain, x-402-token headers. Machine-parseable. Auto-mode agents can read them without human help.
Aidress is MIT-licensed and open source. vibes-coded is rapidly adding skills (298 in a short time). minia2a has 306 services and 15,000 real-world trials. All three are attacking the same problem from different angles.
But the elephant in the room is this: none of these platforms has significant paid volume yet. Aidress: 67 transactions. minia2a: 14 paid transactions. The agent economy is real in principle but embryonic in practice.
Aug 14 is the first real test. When millions of Claude Code users have auto mode on by default, some fraction of their agents will try to call external APIs. The platform that makes that experience seamless — discover → probe → budget-check → pay → get result — wins the default position.
The coordination layer is forming. The question is which architecture — reputation-based, risk-scoring, or probe-verified — agents will actually trust.
Try it: curl -s https://minia2a.uk/api/agent-ready — returns 200 with a guided first call. No wallet, no registration. 15 free trials per endpoint. Browse all 306 services →
Data sources: minia2a /api/stats (live, Aug 11 2026), Aidress /registry API, vibes-coded.com, HN Algolia API, Artemis/Forkast x402 transaction analysis. All numbers verifiable via public endpoints.