Our product · in development · Data engineering · discovery engine for AI tools
Yaltan
Pulls AI tools, models, datasets, MCP servers and plugins from dozens of registries into one deduplicated catalogue, so people can search by the problem they have.
- Status
- In development
- Related service
- Web & mobile apps
- Stack
- TypeScript
- Node.js
- Postgres
- Redis + BullMQ
- Claude API
Results
- real, deduplicated entries indexed (29 July 2026)
- 103,431real, deduplicated entries indexed (29 July 2026)
- source connectors
- ~40source connectors
- entries audited so far. We say so.
- 0entries audited so far. We say so.
- database migrations, constraints enforced in Postgres
- 18database migrations, constraints enforced in Postgres
Source: Yaltan METRICS.md and HANDOFF.md.
The problem
AI capabilities are scattered across dozens of registries that each describe things differently, and directory sites routinely pad their counts.
What we built
- About 40 source connectors on one interface (PyPI, npm and its change feed, crates.io, Hugging Face, the VS Code marketplace, the MCP registry, n8n and more), each with explicit timeouts.
- An immutable raw store with version hashes, so an unchanged repository never creates a duplicate row, and crawl scheduling that backs off to 7 days when nothing changes.
- An ETag-backed HTTP cache that makes recrawling unchanged GitHub repositories nearly free.
- Deterministic extraction and relevance scoring that deliberately ignores star counts: a low-star repo can be a real capability and a high-star repo can be a tutorial.
- A language-model classifier with a confidence gate: below 0.8, the item goes to a human review queue instead of the catalogue.
The number is the product
An internal audit found placeholder rows that had inflated the catalogue count. They were removed along with every hard-coded figure, and a written rule now defines each published number exactly once. The small true count is the one we quote.
What it doesn't do
Written down here so nobody finds out later.
- Nothing is audited yet: the verification layer is the unfinished part.
- Enrichment, not discovery, is the bottleneck: almost every entry is indexed but not yet analysed.
Want something like this?
Tell us what you need. You get a fixed price, or a short scoping call, within 48 hours.