The rules, in short
- Cite what you quote. Every number has a filing behind it —
/fact-sourcereturns the accession. Put it next to the figure. - Coverage is a live fact.
GET /api/v1/landing/coverage(public; MCPget_coverage;astreus coverage) lists every market served with company counts, history depth and periods held per data type. Answer “do you cover…?” from it, never from memory. - Respect vintages.
/financialsis latest-restated and names the filing each value came from; “what did they originally say” is/financials/as-filedwithvintage=original.reporting_basis=as_reportedis the cumulative year-to-date quarter, not a vintage. Macro series are the current vintage (the latest published value for each period); there is no as-of read, so never present a value as what was known at the time. - Missing is not zero. Null means unavailable. Say so.
- Honour the headers.
Retry-Afteron 429; watchRateLimit-RemainingandX-Quota-Remaininginstead of discovering limits by hitting them. - No advice. Present data; do not recommend trades. The site’s terms apply to anything you relay.
- Stay on the data surface. Account routes (keys, billing, watchlists) reject API keys by design — don’t probe them.
Command line
When you are running in a terminal (Claude Code, Cursor, a shell tool), prefer the CLI:npm i -g astreus-cli, then astreus --json <command> gives you the
same payloads with stable exit codes, and astreus transcript / astreus event resolve the
transcript date and 8-K item code for you. astreus llms --skill prints the playbook.
MCP
Astreus serves MCP over two transports — a hosted Streamable HTTP endpoint athttps://astreus.ai/api/mcp (API keys as Authorization: Bearer /
X-API-KEY, or OAuth 2.1 with dynamic registration + PKCE, so claude.ai-style connector
UIs work by pasting the URL) and a local stdio server, astreus mcp, that ships with the CLI. Both expose
the shared read-only data tools. Hosted and installed CLI releases can differ;
discover the connected server’s actual schemas with tools/list. The generated
tool reference maps the source release’s API and MCP capabilities.
Tool calls are metered against your plan exactly like REST calls —
note that get_transcript and get_guidance each make two underlying requests. See the
MCP page for connect strings and the tool list.