Zactonz MCP Server

v0.3.0

A Model Context Protocol server that gives AI assistants and agent fleets the Zactonz APIs as tools, over stdio or HTTP, with structured results: screenshots, page to Markdown, link previews, DNS, SSL, WHOIS, email checks, QR codes, barcodes, social images, image conversion and translation.

Node.js 20+An MCP client over stdio or HTTPZactonz API keys Updated 17 Nov 2026

Security & limits

What is sent and kept#

  • Tool arguments go to api.zactonz.com and nowhere else. The server keeps no logs and stores nothing on disk. The API's own handling of data is covered by the privacy policy.
  • Keys go to the API and nowhere else. They are sent only in the Authorization header to the API and are never included in anything returned to the assistant. Over HTTP, keys a request brings are used for that request alone and discarded with it.
  • Arguments are validated against each tool's schema before any request is made, and arguments outside the schema are refused.
  • Images are fetched for inline display only from the API's own host, over HTTPS, without following redirects, and capped in size while streaming.
  • Tools that fetch a URL do so from Zactonz's servers, which refuse private and internal addresses.

Untrusted content#

read_webpage, convert_html_to_markdown, preview_link and read_qr_code return text written by whoever controls the page or the code. That text can contain instructions aimed at the assistant. The server labels it as third-party data, but the label is advice to the model, not a guarantee. Review what an assistant does after reading pages you do not control.

Screenshots, PDFs, QR codes, barcodes and converted images are stored on api.zactonz.com at unguessable URLs that anyone holding the link can open, for the period given on each endpoint's reference page. Do not render documents that must stay private.

The HTTP transport#

  • It binds loopback and serves no browser. A request carrying an Origin header is refused, as is one whose Host header names a host the server was not given. Those are the two headers a page cannot be made to lie about, and so the two that stop a site you are visiting from reaching a server bound to your own machine. Both refusals are logged with the value and answered without it. The endpoint sends no CORS headers and answers OPTIONS with 405.
  • It has no authentication of its own. A caller is whoever holds a Zactonz key. On loopback that is the machine's own user; anywhere else it is whoever can reach the port, so put TLS and whatever authentication you need in front of it. Keys travel in a header, which is cleartext without TLS.
  • Binding away from loopback is reported on stderr at startup, with what it means.
  • Load is shed, not queued. Requests beyond ZACTONZ_MCP_MAX_REQUESTS at once are refused with 503 and a Retry-After, and a body over ZACTONZ_MCP_MAX_BODY_BYTES is refused rather than read.

Timing and quota#

Calls spend units from your plan's quota, the same as direct API calls. See rate limits and quotas.

A tool call is given 55 seconds in total, because MCP clients stop waiting after 60. Within that time the server retries once, and only where a retry cannot repeat work: after a 429 with a short Retry-After, after a gateway error on a read, or ten seconds after the API host's request burst guard rejects a call. If the client cancels a call, the server stops.

The process sends at most ZACTONZ_MCP_MAX_CONCURRENCY calls to the API at once, eight by default, and queues the rest, counting the wait against the call's own 55 seconds. The firewall in front of the API blocks a burst of more than twenty requests to one endpoint within two seconds, so without a cap a fleet of agents sharing one process would have every call blocked rather than served in turn.

If the API fails to answer ZACTONZ_MCP_BREAKER_FAILURES calls in a row, five by default, the server stops calling it for 30 seconds and answers in the meantime with how long is left. Only the API not answering counts: a refusal, a limit or a bad argument all mean it is there. When the wait is over, one call goes out to find out whether the API is back and the rest wait for its answer, so recovering costs one call rather than all of them.

Reporting a problem#

Bugs and feature requests go to the issue tracker. Report a vulnerability privately by email as described in the repository's security policy; do not open a public issue for it.