Trove Toolkits & Sources

Toolkits and sources that connect Trove (and the agent) to the outside world. There are two kinds — see docs/taxonomy.md:

• Toolkits (mcp/) — named bundles of tools the agent calls directly, live, to read or act. Every toolkit runs as a full MCP server on Trove's cloud, built on the @ontrove/mcp SDK and sandboxed over public APIs. (the current focus — see the catalog)
• Sources (sources/) — fetch content you can already access (your own feeds and accounts, public APIs, RSS/Atom, sitemaps, archives), polling on a cadence and tracking what's new, and fill the searchable knowledge base. Each is a module exporting sync(ctx) → { documents, cursor, stats }.

Why Trove?

AI agents are powerful reasoners but poor data gatherers. Trove gives them two ways to reach real data — toolkits they call live, and sources that fill a searchable knowledge base in the background. On their own, agents struggle to:

• Reach trustworthy data live — query arXiv, SEC EDGAR, FRED, Wikidata, a map API, or the weather through one typed /mcp endpoint, instead of guessing.
• Reach your own accounts and feeds — your newsletters, your read-later list, your podcast queue.
• Maintain state across sessions — knowing what's already been fetched vs. what's new.
• Access historical archives — years of RSS history, paged API backfills.

Each toolkit is a sandboxed server over a public API; each source encapsulates one feed or account's complexity — pagination, rate limiting, incremental resume — behind a simple sync(ctx) → documents interface.

What's Included

Toolkits (mcp/) — called live by the agent

Read-only hosted toolkits over public APIs (plus one mutating toolkit, resend, for send-email) — arXiv, Wikipedia, Wikidata, OpenAlex, Semantic Scholar, Open Library, Internet Archive, Project Gutenberg, SEC EDGAR, FRED, World Bank, Canada Open Data, OpenParliament, Mapbox, Open-Meteo, PubChem, openFDA, The Met, and more. See the full catalog in mcp/README.md.

Sources (sources/) — fill the knowledge base

Filed by subject, Dewey-decimal style (folder name = category):

Folder	What's in it
650-business	Stratechery, Not Boring, AVC, Lenny's Newsletter, Bits about Money, SEC filings
600-technology	Benedict Evans, Simon Willison, Daring Fireball, Benn Stancil
070-news	Guardian / BBC / NYTimes / FT headlines, Hacker News, The Conversation
330-economics	Marginal Revolution
500-science	arXiv papers, Quanta Magazine
300-social-sciences	Our World in Data
000-general	Apple Podcasts, OpenStax, X bookmarks, any RSS feed

Requirements

• Bun ≥ 1.2 — the project's package manager, runtime, and test runner (no Node toolchain required)

Quick Start

bun install

Run a source

The headless runner exercises any source through the same context contract the app uses, printing the resulting documents, cursor, and stats:

# RSS source (no auth, instant)
bun run source sources/070-news/hacker-news --json

Incremental sync

Sources return a cursor that you pass back on the next sync to skip already-fetched content:

const result1 = await sync(ctx);          // First sync: 48 documents
ctx.cursor = result1.cursor;              // Save cursor
const result2 = await sync({ ...ctx, cursor: result1.cursor }); // 0 new, 48 skipped

Building a Source

The simplest source is 8 lines — an RSS feed URL and a prefix:

import { syncRSS } from '../../li