Show HN: Klaus – a Claude Code native delegating agentic harness

Good afternoon everyone, spent like 2 weeks straight doom-scrolling Anthropic docs, Arxiv rabbit holes, and Claude blog posts till my eyes bled. Instead of shipping yet another wrapper that "reinvents" the wheel, I just vibe engineered this thing to max out Claude Code's native features. Klaus Baudelaire (https://github.com/blas0/klaus-baudelaire) basically: Routes agents (in parallel/sequence) with through keyword based scoring + prompt length (no fancy external router needed).

Instead of trying to make agentic delegation deterministic by setting keyword variables as a "YOU MUST INVOKE X-AGENT IF THE USER SAYS DELEGATE", keywords = a score, a certain score = a specific system prompt that is injected into the session post UserPromptSubmit. You can read how this works more on the repo: (https://github.com/blas0/klaus-baudelaire)

Delegates built-in agents + some redefined ones (broad general roles + hyper-specific tool callers) Some agents were built from/around Claude's exposed system prompts and whatever invisible native agents they're hiding behind the scenes.

Bonus: optional memory manager tuned to how Claude actually thinks (not some vector DB cope) to help with managing memory based on: https://code.claude.com/docs/en/memory - doing away with the branded ".memory" directories and such.

Zero external APIs, zero services, zero reinventing (besides context7).

Feels like what Anthropic probably wanted us to do before everyone started LARPing as OpenAI. Plugin version drops late Monday–early Wednesday (this week, Jan 2026).

Video coming day-of with routing/delegation demo + bare Claude Code vs Klaus Baudelaire.

Thoughts? Go ahead and roast or don't! Please, just don't make another 500-line prompt wrapper pls and call it a novel abstraction.