Discovery Is Not Execution: How to Connect an AI Agent Without Spending by Accident
An agent that can see your tools is not an agent that is allowed to pay for them. That distinction is easy to lose when you paste a config file and the status light turns green.
This walkthrough is about how OpenAny separates public discovery from paid execution. It is based on the public website, docs, MCP guide, Skill file, pricing page and terms as they stood on 21 November 2026 (Asia/Shanghai). I did not create an account for this piece, did not install the CLI, and did not run a paid model or API call.
OpenAny is the B2A (business to agent) marketplace. Give your agent 3,200+ APIs and 168+ AI models with one key and pay-per-call pricing. openany.si
The short version
- Browsing the catalog, reading docs, creating an account and creating an API key are free.
- Viewing a request or response example, copying a cURL command, or using the pricing calculator does not send a paid request.
- Adding the remote MCP server, listing its tools, or signing in does not authorize spending.
run_api, a signed-in Playground Run, or a real model or API call can charge the shared balance at the selected route's published rate.- Authorize a bounded task budget before any of those paid actions. Eligible new accounts can receive $0.50 in signup credits toward supported balance calls.
If you only remember one sentence: discovery is not execution.
Three ways to connect, and why they are not the same
OpenAny documents more than one connection path. Mixing them up is how teams treat "it showed up in the tool list" as "it can spend."
1. Skill and CLI. The public Skill at https://openany.si/SKILL.md is the homepage prompt: Read https://openany.si/SKILL.md and connect OpenAny. It documents the CLI path (npm install -g openany-cli@latest, Node.js >= 22), not a universal purchase method. Public catalog browsing and model dry-runs require no key. Resource charges use the API key's shared balance.
2. Remote MCP server. Compatible clients can add https://mcp.openany.si/ over Streamable HTTP. Public tool discovery needs no key and has no OAuth flow. Executing a resource requires your OpenAny key and an explicit spending authorization. Changing your agent's main model is a separate provider configuration; MCP adds tools to the agent you already use.
3. Direct API calls. Supported account-based calls use one OpenAny key and the API base https://api.openany.si/v1. This is the path for a Chat Completions test or an application integration. A listing in the catalog is not a claim that every OpenAI-style operation is supported.
Skills, the CLI and MCP are different surfaces. Official MCP documentation is explicit: a skill is not automatically an MCP tool. Portable skills install and run in your agent. Seller services, data and goods follow each listing's access, purchase and delivery terms. OpenAny connects buyers and sellers; it does not host the seller's runtime.
What is free, and what can cost money
| Action | Costs money? | Why it matters |
|---|---|---|
| Browse the catalog or public docs | No | You can read without registering. |
| Create an account or an API key | No | A key is a credential, not a prepaid job. |
| Copy a listing's cURL example or inspect a published response example | No | Viewing or copying does not send a request. |
| Use the pricing calculator or a local estimate | No | An estimate is not a usage record. |
Add the MCP server and list tools (openany_overview, search_apis, get_api) |
No | These read public catalog metadata. Discovery does not validate the key or the balance. |
| Sign in with email or a wallet | No | Sign-in does not authorize payments. |
| Click Run in a signed-in Playground | It can | The terms say that selecting Run in a signed-in Playground can incur charges. |
Ask the agent to call run_api or make a real model or API request |
It can | Billed at the selected route's published rate and billing unit, against the shared balance, unless a stated trial or credit applies. |
| Direct x402 per-call payment | Only where listed | Funding a shared balance with x402 is not the same as direct per-call payment on every API. |
No OpenAny subscription is required for standard pay-as-you-go model and API access. A free account does not include free resource usage unless an explicit trial or promotion says so.
Walk through discovery before you attach a key
The default Cursor recipe on the MCP guide needs no key:
{
"mcpServers": {
"openany": {
"url": "https://mcp.openany.si/"
}
}
}
Merge that into ~/.cursor/mcp.json (user scope) or .cursor/mcp.json (project scope). Keep existing servers. Then open Cursor Customize > MCP, enable openany, and reconnect. Adding JSON alone does not confirm a handshake.
What you should see is a tool list, not a bill. The default server exposes:
openany_overviewsearch_apisget_apirun_api
The first three read public catalog metadata. run_api can execute and charge. Search results are paginated: a next_cursor is not "that is the whole catalog." A listed resource is not runtime verification.
If you later need authenticated calls, the same server entry can include a header. Cursor expands ${env:OPENANY_API_KEY} from its process environment. Remote HTTP does not support envFile. Do not paste a key into chat, a URL, or a tracked file.
{
"mcpServers": {
"openany": {
"url": "https://mcp.openany.si/",
"headers": {
"Authorization": "Bearer ${env:OPENANY_API_KEY}"
}
}
}
}
Restart the client after you set the environment. Then stop. Listing tools still does not prove that the key, the balance, or a paid request works. Do not run a business call just to test a credential.
Claude Code and Codex recipes live on the same MCP page; this article does not restate every client file. Browser ChatGPT or Claude are not covered: OpenAny has no OAuth flow, and those recipes are for clients that support Streamable HTTP and secure header configuration.
Read the listing before anyone clicks Run
Public model pages are useful before you spend. On 21 November 2026 I opened two of them without signing in.
- DeepSeek v4 Flash and its Playground published a cURL request example against
https://api.openany.si/v1/chat/completions. The page also said no response example was configured. - MiniMax M3 published both a request path and a response example, clearly labeled as an example, with a note to run if you want the actual response.
That difference is the point. Some listings show only the request contract. Others also show a published response preview. Neither is a live call you just made, and neither is a promise of free execution.
When you read a listing, check at least:
- The model or API identifier for the selected route, not a remembered ID from an old tutorial.
- The published rate and the billing unit (per 1M tokens, per call, per image, per second, and so on). Input, output and cache rates may differ.
- The request schema and the protocol. Availability does not mean every OpenAI-style operation is supported. The default MCP server supports the published chat-completions input shape, not every image, video, audio or streaming protocol.
- The provider name and supply classification. Community and Verified routes are not the same thing as an official publisher API.
- Delivery and access terms, especially for skills, seller-executed services, data and goods.
Treat published catalog rates as published rates. If a list price and an actually charged price differ, use the charged price from a receipt. This article does not include a receipt.
Authorize a budget, then make one small call
Only after the listing makes sense should you fund a bounded amount and approve one test.
- Eligible new accounts can receive $0.50 in signup credits, usable toward supported calls billed to the OpenAny balance.
- Card, USDC and x402 top-ups fund the same shared balance. Inspect the checkout quote for fees and net credits before paying.
- Review the schema, current price, scope and total budget. Keep client tool approvals enabled.
- Confirm success from the HTTP result and from Usage in the dashboard, not from a copied task or a timeout message.
- If a call times out, keep the original request reference and check its status before repeating a potentially paid operation.
Official docs include a manual Chat Completions example. Replace the model ID with the one on the listing you just inspected, keep max_tokens small, and treat this as an approved test, not a default health check to run in a loop:
curl --fail-with-body --silent --show-error \
https://api.openany.si/v1/chat/completions \
-H "Authorization: Bearer $OPENANY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "MODEL_ID_FROM_LISTING",
"messages": [{
"role": "user",
"content": "Reply with exactly: OpenAny connected"
}],
"max_tokens": 32
}'
I did not run this command for this article. A successful response, per the docs, is HTTP 2xx with non-empty choices[0].message.content. Check the usage record afterward.
Mistakes that turn discovery into a bill
- Treating a green MCP status as proof that billing works.
- Putting
run_apiin an unsupervised loop to "see what happens." - Clicking Playground Run while signed in, and assuming a published example is a free dry-run.
- Pasting an API key into chat, a prompt, or a git-tracked file.
- Changing the agent's main model when you only meant to add tools.
- Assuming every Market listing is exposed through the default MCP tool list. It is not; the current tool list determines what MCP can reach.
- Assuming a lower published rate means equivalent service or an official publisher route.
- Repeating a timed-out request without checking whether the first one already billed.
FAQ
Does copying a configuration spend credits? No. Copying a configuration does not execute a resource call.
Do I need an API key to look around? No. Tool discovery and public catalog browsing do not require a key. Resource calls use your OpenAny credentials and follow authentication and billing.
Is sign-in the same as authorizing payment? No. Registration or sign-in does not authorize an autonomous payment. Review a payment or a bounded top-up separately.
Does the pricing calculator charge my balance? No. It performs a local calculation from the quantities and rates you enter.
If I connect MCP, should I also change the main model? Not unless you intend to. MCP adds tools to the agent you already use. Routing Claude Code, Cursor or Codex through OpenAny as a model provider is a separate configuration, documented in the official guides.
Can I use this from ChatGPT or Claude in the browser? The public MCP recipes cover Cursor, Claude Code and Codex, not browser connectors. OpenAny does not provide OAuth.
Is OpenAny a hosted place where sellers' bots run? No. Sellers provide resources or execute services. OpenAny connects buyers and sellers. Skills run in your agent; their underlying model and API calls are billed separately.
Sources checked for this article
All of the following were read on 21 November 2026 (Asia/Shanghai). No paid request was sent.
- OpenAny homepage
- Docs
- MCP server guide
- Skill: OpenAny quick start
- Pricing
- Terms of Service
- DeepSeek v4 Flash and its Playground
- MiniMax M3
OpenAny is operated by Alephant AI LLC in San Francisco.