what the lab builds
Everything the lab builds, flagship first: KVWarden is live and measured, mlxd and Coconut OS are in research, and the smaller tools behind the public work sit at the end.
One lab, several surfaces. kvwarden.org and coconutos.org are separate domains for now, but everything here is Coconut Labs.
Tenant fairness on shared inference.
1.14× of solo TTFT, 26× better than FIFO
KVWarden is a scheduler and cache-pressure experiment for shared LLM inference. The first public result is narrow on purpose: a quiet tenant stays near solo latency while a flooder pushes the system. The harness is public; the plots do not hide the quiet tenant in an aggregate.
Tenant-fair LLM inference on Apple Silicon.
mlxd is a planned scheduler and admission layer that lives on top of `mlx_lm.server`. The thesis: today MLX has no tenant identity, so concurrent requests can bleed KV cache between callers. Once correctness is restored, the next gap is fairness, and the model has to be different from CUDA's KV-block partitioning because unified memory makes bandwidth the shared resource, not GPU memory. Sibling to KVWarden under coconut-labs. Shared methodology, separate codebase.
A Linux distribution where agents are first-class.
Coconut OS is the lab's largest bet: a Linux distribution built on the position that agents should be kernel primitives, not processes with extra steps. Today an agent on Linux is indistinguishable from any other process, so everything about it lives in userland convention: what it may touch, what it did, who let it. We think the kernel should know. The work so far is specification and prototype, and we publish the posture honestly: a capability subsystem with explicit tokens for agent operations, an audit path that records what agents did at the kernel boundary rather than trusting them to report it, and scheduling that knows an agent workload from an interactive one. The reserved syscall surface is small on purpose, and most of it is deliberately unwired until the design survives review. Same lab, same method as KVWarden and mlxd: state the thesis, build the smallest honest version, measure, publish what happened. The landing page lives at [coconutos.org](https://coconutos.org), and like every domain we run, it is one surface of one lab. ## The specs, in full Five specification pages are public and had no inbound link from this domain until now. In reading order: - [Thesis](https://coconutos.org/specs/thesis), the position the whole distribution rests on. - [Architecture](https://coconutos.org/specs/architecture), the capability subsystem, the audit path, and the syscall surface. - [Security](https://coconutos.org/specs/security), the threat model and what the capability tokens are supposed to stop. - [Roadmap](https://coconutos.org/specs/roadmap), what is wired today and what is deliberately unwired. - [Team](https://coconutos.org/specs/team), who is doing the work.
an agentic MLOps atlas + measured prototypes
One method at two scales: a deep agentic-MLOps platform case study, and seven small free-standing prototypes, each solving one company’s real data-systems bottleneck, with a guardrail you can run in your browser.
tools and experiments
Smaller things, mostly the scaffolding behind the public work.
Small scripts, traces, and notebook fragments behind the public research feed.