io.github.EvidInvest/aether-developer

aether-developer

Everything you need to connect Aether — the search engine built for agents, over SEC filings, earnings-call transcripts, and EU financial regulation — to your LLM, your IDE, or your own code.

Every result is a ready-to-cite evidence payload: the exact filed passage, an accession-numbered citation, the sec.gov / EUR-Lex link, and a confidence score. No filing, no answer — your agent stops inventing numbers.

aether-developer/
├── docs/                — connection guides + REST API reference
│   ├── chatgpt.md       — ChatGPT Developer Mode setup
│   ├── mcp.md           — Claude / Cursor / Cline / any MCP client
│   ├── cursor.md        — Cursor specifics (one-click install)
│   └── search.md        — REST API reference (/v1/tools/*), the issuer contract, fetch mode
├── skills/              — Agent Skills (drop into ~/.claude/skills)
├── mcp/                 — @evidinvest/aether-mcp (stdio MCP server, npm)
├── clients/
│   ├── typescript/      — @evidinvest/aether-sdk (npm)
│   └── python/          — evidinvest-aether-sdk (PyPI)
└── examples/            — runnable demos + config snippets

Connect your LLM (no code)

MCP endpoint: https://api.aether.evidinvest.com/mcp · Auth: OAuth (sign in when your client prompts — free account).

Then ask: "What does Apple's latest 10-K say about supply-chain risk? Quote the passage with the sec.gov link." Name the company in the question — every tool takes an issuer, and naming it is what makes the answer scoped rather than a relevance-ranked sweep of every filer.

Teach your agent the craft (skills)

MCP gives the agent the tools; the skills/ directory teaches it the craft — tool routing, citation discipline, point-in-time transcript retrieval:

cp -r skills/aether-research ~/.claude/skills/   # Claude Code, personal

See skills/README.md for Claude Desktop / other frameworks.

Call the API from code

Four REST endpoints — the same tools MCP exposes (full reference):

POST /v1/tools/search   ← start here (unified)
POST /v1/tools/financial_search    POST /v1/tools/transcript_search    POST /v1/tools/regulation_search

Name the company, then narrow. Pass issuer on every call and the answer is scoped to that filer: fast and precise. Omit it and the call searches every issuer at once — slower, relevance-ranked only, and the response carries a quality_caveat. scope: "issuer" with no caveat is the only guarantee you got a filtered answer, so read both fields.

And query is optional: omit it, pass an identifier plus form_type / fiscal_year / section, and you get fetch mode — that filing's sections in filing order, no ranking, mode: "fetch", tens of milliseconds. If you already know which document you want, don't rank it.

TypeScript / Node

pnpm add @evidinvest/aether-sdk

import { AetherClient } from "@evidinvest/aether-sdk";

const aether = new AetherClient({ apiKey: process.env.AETHER_API_KEY });

// search: the question in `query`, the company in `issuer`
const out = await aether.financialSearch({
  query: "supply-chain concentration risk",
  issuer: { ticker: "AAPL" },
  fiscal_year: 2025,
  limit: 5,
});
if (out.quality_caveat) console.warn(out.scope, out.quality_caveat);
for (const c of out.results) {
  const start = c.metadata?.retrieval?.body_offset ?? 0; // skip the EDGAR cover page
  console.log(c.citation, c.metadata?.source_url, c.text.slice(start, start + 200));
}

// fetch: no query — Apple's latest 10-K risk factors, in filing order
const risks = await aether.financialSearch({
  issuer: { ticker: "AAPL" },
  form_type: ["10-K"],
  section: "Item 1A",
  limit: 5,
});
console.log(risks.mode); // "fetch" — and no `confidence` on any hit

Python

pip install evidinvest-aether-sdk

from aether import AetherClient

with AetherClient(api_key="ak_...") as aether:
    out = aether.financial_search(
        query="supply-chain concentration risk",
        issuer={"ticker": "AAPL"},
        fiscal_year=2025,
        limit=5,
    )
    if out.quality_caveat:
        print(out.scope, out.quality_caveat)
    for c in out.results:
        start = c.body_offset or 0          # skip the EDGAR cover page
        print(c.citation, c.metadata.get("source_url"), c.text[start : start + 200])

    # fetch: no query — Apple's latest 10-K risk factors, in filing order
    risks = aether.financial_search(
        issuer={"ticker": "AAPL"}, form_type=["10-K"], section="Item 1A", limit=5
    )
    print(risks.mode)   # "fetch" — and `confidence` is None on every hit

curl / any language

curl -sS https://api.aether.evidinvest.com/v1/tools/search \
  -H "authorization: Bearer $AETHER_API_KEY" \
  -H "content-type: application/json" \
  -d '{"query":"supply chain risk Taiwan","issuer":{"ticker":"TSM"},"limit":5}' | jq .

Docs

• docs/search.md — REST API reference: all three tools, params, response shape.
• docs/regulation_search.md — EU financial-regulation search deep-dive (29 acts), filters, curl.
• docs/mcp.md — connect any MCP client (remote URL or stdio), env overrides.
• docs/chatgpt.md — ChatGPT Developer Mode walkthrough.
• docs/cursor.md — Cursor one-click install.

Get an API key

https://aether.evidinvest.com/developer/keys — anonymous calls work but are rate-limited; keys are free.

What's in the corpus

• SEC filings — 10-K / 10-Q / 8-K, registration statements, prospectuses, press exhibits; ~10 years of S&P 500 and beyond.
• Non-US registries — Sweden (Bolagsverket), Japan (EDINET), Korea (DART) annual reports. Reach a non-US filer by name — issuer: {company_name: "Sivers Semiconductors"} — or scope a whole market with jurisdiction: ["SE"|"JP"|"KR"].
• Earnings calls — speaker-attributed transcripts + furnished press exhibits, with point-in-time filters (order: "earliest" finds first mentions) and the issuer's own fiscal_year + quarter.
• EU regulation — MiFID II, MiCA, DORA, the AML package; article-level citations.

3 months free. Verify your email when you sign up and you're on the trial tier automatically — up to 5,000 API calls/hour on Aether's search tools, free for 3 months. No card required; it reverts to the free tier (100 calls/hour + paid credits) when the trial ends.

Contribute a client library

The TypeScript and Python clients are intentionally tiny — one bearer-token fetch wrapper + typed shapes for the /v1/tools/* endpoints. Porting to Go, Rust, Java, etc. is ~150 lines. Open a PR under clients/<lang>/. Whatever you write, surface scope, quality_caveat, mode and each hit's body_offset: a client that drops them leaves its users unable to tell a precise answer from an approximate one. Response fixtures to test against live in clients/fixtures/.

License

Apache-2.0 — see LICENSE.