Layagrep is a local code retrieval CLI for coding agents. It forks jevgrep (MIT) and replaces its hosted relevance decisions with Laya running on your machine. Ask what code does; Layagrep returns relevant files, source excerpts, and line references. Source reads stay local when the Laya server is bound to loopback.
A separate project, marksomething/layagrep, uses the same name for semantic line matching. This fork focuses on repository discovery and source context.
Requires Node.js 22+, Bun for building this checkout, Python 3.10+, and enough memory for a Laya checkpoint. A CUDA GPU is optional.
uv tool install 'laya[serve]'
LAYA_HOST=127.0.0.1 LAYA_DEVICE=cuda LAYA_PRELOAD=1 laya-serveIn another terminal:
bun install
bun run --cwd apps/cli build
npm install --global ./apps/cli
layagrep doctor
layagrep "Where is authentication checked before a request reaches a handler?" /path/to/repoThe default endpoint is http://127.0.0.1:8000/v1. If port 8000 is occupied, set LAYAGREP_LAYA_URL=http://127.0.0.1:8765/v1 or save that URL in ~/.config/layagrep/server-url; only loopback HTTP endpoints are accepted. No external API key is required. If you set LAYA_API_KEY on the Laya server, run layagrep auth to store the matching key locally. layagrep skill installs the bundled agent skill. jg remains an alias for existing scripts.
Laya's multilingual checkpoint is selected explicitly, with a 4096-token request budget. This fork has not reproduced jevgrep's published cost or retrieval benchmarks; those upstream measurements do not establish Layagrep's quality. If a search reports incomplete discovery, inspect the source directly. Exact symbols and paths are often faster with rg.
We ran GPT-6 Sol once on each of the same ten public SWE-bench tasks selected for the jevgrep comparison with and without the Layagrep skill. Both variants used the same pinned task images and official SWE-bench grader. The skill used a local Laya server; CodeGraph was not enabled.
| Without Layagrep | With Layagrep | |
|---|---|---|
| Officially resolved | 7/10 | 7/10 |
| Total agent time | 20 min 19 s | 27 min 18 s |
| Estimated GPT-6 Sol API cost | $2.31 | $2.63 |
Layagrep produced no additional solves in this run. The skill runs took about 34% longer and used about 14% more model API spend. Agent time includes search, editing, and checks, but excludes image downloads and official grading; costs use standard GPT-6 Sol token prices and exclude local Laya compute. One task (pylint-dev__pylint-4604) was inconclusive in both arms because the grader collected no tests. An initial skill setup failure was excluded from the paired results and cost an additional estimated $0.12 in API usage.
You only know whether a tool helps when you test it. A trial that finds no improvement is useful evidence too. This is one attempt per task with one model and no CodeGraph index, so it does not establish how Layagrep performs on other repositories or configurations.
The CLI can use CodeGraph as an optional local structural index. CodeGraph has its own file watcher and incremental sync; Layagrep does not maintain a second graph database. Install CodeGraph with npm install -g @colbymchenry/codegraph, run codegraph telemetry off if you want usage telemetry disabled, and then run codegraph init /path/to/repo. Search with layagrep "question" /path/to/repo --graph. Layagrep runs CodeGraph's incremental sync before each graph search, so edits are picked up even if no watcher is active. CodeGraph's serve --mcp mode additionally watches files continuously. Graph results scope the Laya search; rerun without --graph when you need whole-tree discovery. Layagrep checks graph paths against the current files and its normal filesystem policy before returning excerpts.
Layagrep is MIT licensed and retains jevgrep's copyright and history. Laya is Apache-2.0 licensed and is installed and run separately; no Laya source, model, or binary is distributed with this repository or the Layagrep CLI. CodeGraph is also a separately installed tool. If a future distribution bundles Laya, it must include Laya's Apache-2.0 license and any applicable notices with that distribution.