What is running, and what it costs.
Radical transparency about how the desk runs. This page is itself static and served from the edge. It documents the machinery, its running cost, and the endpoint that reports the live truth.
The moving parts, and how they read.
The labels below identify documented parts of the architecture. They are not live status badges. A static export cannot poll its own uptime and then honestly print it, so it does not pretend to. For the current reading, the desk runs its own pipeline monitor at /api/health. Call it yourself: it re-runs the load-bearing upstream legs and reports each one, returning status: "ok" only when the news wire and the quote tape both pass.
The pages you are reading. A Next.js 16 static export, pre-rendered at build time and served from Cloudflare Pages. No request-time rendering, no origin to go down.
/api/* Cloudflare Pages Functions hold the keys and answer in JSON. Successful GET responses are edge-cached by endpoint policy, which reduces repeated upstream work. Caching is not client rate limiting or abuse protection.
/api/news aggregates public RSS feeds, dedupes, and clusters related headlines. It is one of the two legs the /api/health probe treats as load-bearing: an empty wire flips the overall status to degraded.
/api/quote resolves delayed index and equity quotes. Its health leg is a single index resolve; a price that will not parse is the other condition that marks the desk degraded.
/api/wire drafts the daily brief through a chain of free OpenRouter models. When every model is unreachable it composes deterministically from the same data and still returns 200, so this leg is reported by the health probe but never flips the overall status.
/mcp speaks the Model Context Protocol over stateless Streamable HTTP, no key and no login, exposing the desk’s data to an agent. Same terms and the same edge cache as the JSON API.
Near zero, and honest about why.
The desk runs on free tiers, end to end. The site sits on Cloudflare Pages’ free plan; the data comes from free or keyless public sources (FRED, the CFTC, SEC EDGAR, the U.S. Treasury, Yahoo, Finnhub, the EIA); the daily wire is drafted by free OpenRouter models (the :free slugs) with a deterministic fallback when they are out of credits. The marginal cost of one more visitor is effectively $0.
That is a deliberate claim, not a rounded-down one: there is no dollar figure to quote because the running pieces do not bill per request. The per-feed cadence and cost table lives on /methodology, which puts it plainly: everything on the site runs on free or keyless public data.
Fast to serve, honestly cached.
Static pages are served from Cloudflare’s edge, close to the reader, which is fast. There is no measured latency number quoted here, because a static page cannot verify one about itself. What is exact is the freshness contract: each API endpoint carries an edge-cache TTL, and those windows are the honest refresh rates, straight from the OpenAPI document.
Beyond the cache, freshness has one more marker: the build stamp in the site footer records when the current pages were deployed ("Static page · built … ET"). A static page can outlive its deploy by days, so that line is labelled as the build moment, never as data freshness. The full endpoint reference is on /api-docs.
Local fetched data, under a fixed budget.
The desk may retain reproducible fetched responses in this browser to make return visits faster and more resilient. That cache has a fixed 3 MiB payload budget and is separate from the state you create or save.
Only reproducible fetched responses are counted and cleared. Portfolio, watchlists, alerts, preferences, Desk Context, and guided board state are stored separately and never cleared by this control.
Best effort, no SLA.
None of this is a service-level agreement. The endpoints run the live site first and your scripts second. There is no published request-rate limit or uptime guarantee. The cache reduces repeated identical upstream work but does not throttle clients or unique cache misses. It is best effort, disclosed plainly. How the desk holds itself honest is on /standards, and how to call the machine yourself is on /api-docs.