# Operations and privacy

Handle failed or incomplete reads explicitly and keep personal information out of documentation and telemetry.

## Errors and retries

- Check both HTTP status and the MCP JSON-RPC/tool result.
- HTTP 400 means input needs correction.
- HTTP 404 on public content means no published item is available.
- HTTP 429 requires waiting for Retry-After.
- HTTP 503 or `isError: true` means a lookup did not complete. Do not fabricate results or repeatedly retry an expensive request.

## Rate limits

Content, MCP, and API-host limits are independent. Use the returned rate-limit and Retry-After headers as the authority for your connection.

Production MCP requests share a 60-request/minute client limit, a 120-unit/minute client work budget and a 1,200-unit/minute global work budget across instances. Lookups cost one unit, directory searches five, and comparisons or compatibility calls ten. At most 24 MCP requests hold active work leases. Separate local and edge limits also apply.

Public-content reads share a 300-unit/minute global budget and four active work leases; each read costs ten units, with a 120-unit/minute client budget. Publication is checked on every admitted read; responses are not cached. Content requests have a 4 KiB input bound; the MCP content client allows ten seconds and a 1 MiB response bound. Direct public API research has its own shared work budget.

Capacity limits return HTTP 429 with Retry-After. An unavailable shared limiter returns HTTP 503; requests do not continue without accounting. Abandoned leases expire after 90 seconds. These controls reduce abuse; they do not guarantee uninterrupted service during an attack.

## Privacy

Never send account credentials, enrollment identifiers, contact details, or personal health histories to the public tools. Research tools accept the minimum supported inputs for their documented task. Content search accepts general topic keywords only.

Content queries are not logged or sent to analytics. Documentation search runs locally in the browser without transmitting query text. The optional quickstart tag counts only aggregate example outcomes, not unique people. Server access logs classify a fixed list of AI referral hosts without retaining full referral URLs or addresses. This site has no account system, session recording, or API-key portal.

## Source scope

Public docs and editorial content are intended to be discoverable and citable. Plan-data access remains subject to the [published terms](https://thepocketprotector.com/terms), including restrictions on bulk copying. Crawler policy is published in robots.txt. See the [privacy policy](https://thepocketprotector.com/privacy).

## Changelog

This release introduces generated MCP documentation, published-content search and retrieval, and newsroom archive/feed discovery. The current catalog fingerprint is included in the machine-readable catalog. Production activation and domain availability are verified during release; this documentation does not imply that an unreleased tool is already deployed.

## Support

Email [api@thepocketprotector.com](mailto:api@thepocketprotector.com) with the tool name, approximate time, HTTP status and non-sensitive error category. Do not include beneficiary inputs, tokens or raw request bodies.