Splinter - A Breadboard For Complex Vector Workloads

A Cooperative Userspace Hypervisor for Inference & Other Semantic Workloads.

Local LLM inference chokes on the "Socket and Lock" tax — the context switches, serialization, and kernel interrupts that standard IPC tools and databases demand just to synchronize state. At token speeds, that overhead isn't a bottleneck, it's a wall.

Splinter dismantles that wall. It is a lock-free, cooperative userspace hypervisor built for strict mechanical sympathy with modern CPU cache hierarchies (x86_64, ARM, and RISC), putting your governance, your vector storage, and your inference engine in the exact same physical memory lane. Think of it as a semantic breadboard: a passive, shared-memory manifold where thousands of context or classification windows run simultaneously with 100% non-blocking throughput. There is no central arbiter, no socket, and no copy — reads and writes happen at L3 cache speeds, directly on the hardware bus.

If this is your first stop, you're welcome here. Splinter is open source, actively researched, and built in the open. Whatever brought you — embedding the library, driving the CLI, wiring up another language, or just understanding how a lock-free shared-memory substrate holds together — start with the section that fits.

Reference Documentation

  • API Reference — every public function in splinter.h, grouped by purpose: store lifecycle, key/value operations, the seqlock epoch protocol, Bloom labels, the signal arena and event bus, and the Logic Shard election that governs cooperative madvise.
  • CLI Reference — every command for splinter_cli (the interactive REPL) and splinterctl (one-shot shell invocation), from get/set to shard, ingest, and semantic search.
  • Language Bindings (FFI) — Splinter is a very dlopen()-friendly library, so any runtime that can read a Linux DSO joins the same manifold as a first-class peer: Rust, TypeScript (Deno / Bun), Python, Java, Lua, and even POSIX shell.
  • Environment Variables — runtime knobs honored by both the library and the CLI, including SPLINTER_DEFAULT_UMASK (permissions of newly created stores) and SPLINTER_NS_PREFIX (CLI key namespacing).

The Design, In Depth

Every architectural decision in Splinter — the ringless privilege topology, the seqlock that lets readers never block writers, the cooperative bid table that replaces preemptive interrupts, the mop modes that keep stale bytes out of raw reads — is explained in detail, with its rationale, in the thesis:

Collaborating

Contributions of every kind are welcome — code, docs, benchmarks, bindings, bug reports, and grounded "oddball" ideas. Please read CODE_OF_CONDUCT.md and CONTRIBUTING.md before opening a pull request. The maintainer is a brain-cancer survivor doing active research, so pull requests may surge forward and then go quiet for a while; that's normal here, and GitHub issues and pull requests are the way in.

How to Cite

If Splinter is useful in your work or research, please cite it:

Splinter: A Cooperative Userspace Hypervisor for Inference & Other Semantic Workloads.
Tim Post. https://splinterhq.github.io

Attribution and the latest documentation live at splinterhq.github.io.