Show HN: Ratel, give agents unlimited tools and skills without context bloat

AI tool chaos or genius shortcut? Commenters are already calling the bluff

TLDR: Ratel says it can make AI assistants cheaper and less messy by only giving them the few tools they need in the moment. Commenters immediately challenged the hype, with the big debate being whether this is genuinely useful or just old search ideas dressed up in new branding.

A new Show HN launch is pitching a very relatable promise: stop cramming every possible helper, instruction, and memory into an AI’s brain every single turn, and only show it the few things it actually needs right now. The project, called Ratel, says that means cheaper runs, less confusion, and better performance for AI assistants with giant toolboxes. In plain English: instead of dumping the whole kitchen drawer on the counter, it hands over just the spatula.

But the real action is in the peanut gallery, where the immediate reaction was basically: "Wait, isn’t this just the old thing with a shiny new label?" One commenter cut straight to the chase by asking if this is just "RAG for tools"—meaning, is Ratel really a fresh idea, or just a repackaged way of searching through stuff? They also poked at another sore spot: if MCP already has tool search, why do we need another middleman? That one comment pretty much sets the tone of the thread: intrigued, skeptical, and absolutely unwilling to let startup-style branding slide without a fight.

Then there’s the wonderfully awkward side character energy of the other visible reply, which is literally marked as a stub for offtopicness. That accidental anti-joke gives the whole discussion a classic Hacker News flavor: one person questioning the entire premise, another seemingly wandering in from a completely different conversation. So while Ratel is selling order, efficiency, and fewer wasted tokens, the crowd reaction so far is more like: prove it, explain it simply, and don’t pretend we haven’t seen this movie before.

Key Points

  • Ratel is presented as an in-process context engineering platform that selects only relevant tools and skills for each agent turn instead of loading the full catalog into context.
  • The current product focus is tool selection through a ToolCatalog that can register tools directly or ingest an upstream MCP server.
  • The stack includes a Rust core, TypeScript SDK, Python SDK, and CLI, and the article says it requires no vector database, embedding pipeline, or separate deployed service.
  • Ratel uses BM25-based retrieval over schema-aware tool text and a replace-by-default top-K tool injection approach to reduce context size and token usage.
  • The project spans three repositories: the main library, an MCP showcase product, and a benchmark harness used to report model accuracy, token, and cost comparisons.

Hottest takes

"just RAG for tools" — vinci00
"MCP already has tool search" — vinci00
"stub for offtopicness" — tomhow
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.