halowerk-mcp
Provides access to 70 paid APIs as MCP tools with per-call and session budgets, paying via x402/USDC on Base without requiring an account or login. Includes a built-in allowlist of hosts and safe defaults to prevent unintended payments.
README
halowerk-mcp
70 paid APIs as MCP tools. No account, no login, no subscription — a throwaway wallet with a few dollars is enough.
Your spend is capped before anything is signed:
| Variable | Default | What it does |
|---|---|---|
HALOWERK_MAX_PREIS_USDC |
0.05 |
Ceiling per call. A more expensive service is refused, not silently bought. |
HALOWERK_BUDGET_USDC |
1.00 |
Ceiling per session. Once reached, the server refuses further calls. |
HALOWERK_TROCKEN |
off | Dry run: returns the price and the call plan, pays nothing. |
Every paid answer carries bezahlt_usdc, budget_rest_usdc and the transaction hash, so the model and you see the same numbers.
Try it without a wallet
Two tools work with no wallet and cost nothing: halowerk_katalog lists every service with its price, and halowerk_preis tells you what a planned call would cost. Install it, browse, decide later.
{
"mcpServers": {
"halowerk": { "command": "npx", "args": ["-y", "halowerk-mcp"] }
}
}
That goes in your Claude Desktop config (claude_desktop_config.json), Cursor's mcp.json, or Cline's MCP settings.
Paying for calls
{
"mcpServers": {
"halowerk": {
"command": "npx",
"args": ["-y", "halowerk-mcp"],
"env": {
"HALOWERK_PRIVATE_KEY": "0x…",
"HALOWERK_MAX_PREIS_USDC": "0.05",
"HALOWERK_BUDGET_USDC": "1.00"
}
}
}
}
Payment is x402 with USDC on Base. There is no account to create, no card, no invoice: the call carries a signed authorisation for the exact amount, and the service settles it.
Use a throwaway wallet
Create a fresh wallet, put a few dollars of USDC on it, use it for nothing else.
Not because this package is untrustworthy — because you have no reason to trust it yet. A key in an MCP config is a key that a model-driven process can reach. The way to make that safe is not to trust the code but to make the blast radius small: if the wallet holds three dollars, three dollars is the worst case.
You do not need ETH for gas. Payment uses EIP-3009 transferWithAuthorization, so a third party submits the transaction. A balance of exactly 0 ETH is normal and correct.
What the code does with your key
- It is read from
HALOWERK_PRIVATE_KEYand from nowhere else — no file, no argument, no prompt. - It is never logged, never put in an error message, never returned in an answer, not even truncated. There is a test for this:
npm run pruefesearches the entire output for the first eight characters of a key and fails if it finds them. src/x402.tsis about 300 lines and is the only file that touches it. It is meant to be read in five minutes.
Where your money can go
The server only ever pays the ten hosts baked into the catalogue at build time. The list is frozen in the package; it cannot be extended by a file, an environment variable or a tool argument.
This is the point of the whole design. A tool that accepts an arbitrary URL and pays for it is a payment instruction with a blank payee field. A model that reads a web page, a PDF or an email can be told, by that text, to call such a tool with an attacker's address. So:
- The target host must match the built-in list exactly. No suffix comparison —
halowerk.com.attacker.examplepasses anyendsWithcheck and is a different host. httpsonly. Redirects are not followed: a 301 to a foreign host is the attack, not an inconvenience.- No IP addresses, no
localhost, no private ranges. - On violation: refusal with the reason, before anything is signed.
halowerk_rufe_auf takes a tool name from the catalogue, never a URL.
Modes
Seventy tools at once makes the tool picker worse for everybody, especially if you have other MCP servers connected. So:
HALOWERK_MODUS |
Tools | For |
|---|---|---|
hero (default) |
12 services + the 2 free ones | Everyday use |
gateway |
3 meta-tools + the 2 free ones | Minimal context; two extra round trips |
all |
70 services + the 2 free ones | You know what you want |
In gateway mode: halowerk_suche finds a service by description, halowerk_beschreibe returns its input schema, halowerk_rufe_auf executes it.
Why these twelve
One measure only: what does an agent need more than once in the same task? A tool used once per session can be looked up; one used ten times belongs in the toolbox. All twelve are also text in, text out — in a chat window, a tool that needs a file upload first is close to unusable.
| Tool | Repeats because |
|---|---|
cost_estimate |
before every model call |
output_validate |
after every model answer |
tool_arg_validate |
before every expensive or irreversible call |
state_store |
across every step of a long task |
cve_check |
per dependency list, and you look at it more than once |
license_check |
same audit, per package set |
tls_chain |
per host in an infrastructure check |
jwt_verify |
per token while debugging auth |
webhook_signature |
per webhook while debugging an integration |
doc_diff |
per file pair in a review |
web_archive |
per URL in a research task |
invoice_audit |
per invoice in a batch |
Tiered prices — read this before you set a budget
Some services bill per unit: per audio minute, per 1000 characters, per sample point. For those, price_usdc is only the minimum.
The budget check therefore uses max_usdc, the ceiling. Checking against the minimum would approve 0.006 and pay 0.25 — the endpoint bills by length, and the length is in the request body, not in the catalogue.
If a service has no ceiling at all, it counts as unbounded and is not called without unbegrenzt_erlauben: true. A budget that checks against an unknown amount is not a budget.
What this package is not
- It is not a wallet manager. It signs payments; it never moves funds anywhere else.
- It does not create or store keys.
- It does not phone home. The only outbound requests go to the ten catalogue hosts.
- The catalogue is generated from the live services and shipped with the package. It is not fetched at runtime, so a compromised server cannot add itself to your allowlist between calls.
Development
npm run katalog # regenerate katalog.json from the live services
npm run build # compile TypeScript
npm run pruefe # modes, spending brakes, attack tests, key-leak test
katalog.json is generated. Never edit it by hand — regenerate and republish, otherwise the package promises a price the service does not charge.
License
MIT. See LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.