Laya in coding tools
Coding agents make many cheap judgement calls: which files matter, how hard a task is, whether a command is risky. Several projects hand those calls to a local Laya model instead of a second LLM call, which saves tokens and keeps code on the machine.
0 projects
Claude Code
laya-codex indexes a repository locally and gives Claude Code the relevant code before each task, using tree-sitter, BM25 and a Laya re-ranker. Install (macOS Apple Silicon or Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/pilotspace/laya-codex/main/install.sh | sh
/plugin marketplace add pilotspace/laya-codex
/plugin install laya-codex@laya-codex
Its README reports that Claude Code read 38% fewer code tokens and took 21% fewer turns in a paired benchmark of 60 tasks on three repositories.
arbiter serves Laya behind a Jev-compatible API with an MCP server and a Claude Code plugin.
Model routing in coding agents
pi-pignon asks a local laya-serve (or Jev) how hard each prompt is and switches the pi agent's model:
pi install npm:pi-pignon
uv tool install "laya[serve]"
LAYA_HOST=127.0.0.1 laya-serve
Judges and gates
- adecider: typed judgements for any coding agent through a CLI, MCP tool or HTTP, with local Laya as the default backend.
- omp-laya-judge: a local judge for oh-my-pi. It reports a mean 160 ms per question on CPU and 8/12 correct on its benchmark.
- laya-rust: its
exfil_triageexample flags shell commands that send credentials off the machine.
Workspaces
keel is a local-first macOS coding workspace. A local Laya selector (Core ML) or an opt-in hosted Jev picks a route for each new task, and the host validates the choice before applying it.
Caveats
omp-laya-judge found Laya confidently wrong on arithmetic and parity questions (confidence 0.80 and 0.84), so it routes those to the LLM whatever the confidence. Keep Laya on semantic questions and keep a fallback. laya-codex downloads a model of about 850 MB on install.
More ways to use Laya