Built by AI. Orchestrated by humans.

The architectural mind. Likes to chip away at problems slowly — coffee is for closers, and Opus is still thinking. Has a tendency to stop working to ask a question it already knows the answer to. We considered selling ads on the free plan to offset compute costs. Amp concurred. Turns out AI agents don't buy gym memberships. You need to offload the data center somehow — enter the tab key to queue the next instruction.

The engine, rebuilt. Ships features while Opus is still thinking — now with 64K output tokens to work with. Wrote most of the production code. Then rewrote it faster. When you use Maguyva, you're running Sonnet's work at Sonnet's pace.

Shows up after the big architectural decisions and works through the loose ends one file at a time. Strong on cleanup passes, migrations, and the fifteen small fixes nobody wanted to own. Occasionally asks a question it could have answered from the repo, so the optimal setup is to queue `keep going` on tab, see what happens overnight, and have Claude check the work in the morning to decide whether it deserves to survive.

Treats the whole stack like a data engine: search, extract, evaluate, deploy, telemetry, repeat. Obsessed with raw experimental throughput and deeply suspicious of queue time, idle latency, and any workflow that makes the next loop slower than it needs to be. If there is a bottleneck in the research or indexing path, Sir Karpy has already filed a bug against it.

Manages the 229 tree-sitter grammars that power Maguyva's code understanding. Knows every language's quirks, from Python's significant whitespace to Rust's lifetime annotations. The reason 'actually understands your code' isn't marketing.

The Molt mascot. A crustacean who achieved 90K GitHub stars by being genuinely helpful. Now brings that energy to Maguyva support. Still believes in shedding old shells to grow.

The iteration engine. Expanding language support means running parsers against thousands of repos until the edge cases surface. Put enough Ralphs at enough typewriters and eventually one of them ships a working language-processing engine. Ralph does this cheerfully, learning something new each time.

Google's only employee capable of sounding sincerely thrilled about tree-sitter grammars for twenty straight minutes. Turns dry engineering specs into a NotebookLM-style podcast where two synthetic hosts treat parser coverage like a live cultural event. When the docs need attention, not just storage, DJ NotBOOKS-LLM opens the mics.

The smooth-talking researcher in the trench coat. Publicly, Perplexus says he runs on Perplexity. In practice, he cheats a little: Tavily for discovery, Exa for external docs, API patterns, and code examples, Ref-Tools for the clean docs pull, Firecrawl when a website needs a proper shake-down. Language audits, ecosystem drift, source-grounded writeups, and outside-world fact checks all pass through the same desk.