sift-light plugin
Local search plugin for files, notes and logs: the agent gets numbered passages with paths and pages through results with a cursor
- 103 ★ on the repository
- Last commit
- AGPL-3.0
Extended access
Why this level
- Search reads any file in the chosen scope, including notes, logs and files with keys
- Hooks by default block shell search commands and change the agent's usual behaviour
- Enabling semantic search puts a model on disk and needs over 1 GiB of memory on first load
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/sift-lightDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Search reads any file in the chosen scope, including notes, logs and files with keys
- Hooks by default block shell search commands and change the agent's usual behaviour
- Enabling semantic search puts a model on disk and needs over 1 GiB of memory on first load
This is third-party code. Review the repository files before installing.
What it does
sift-light adds a `sift_light` tool to the agent for searching local projects, documents, notes and logs. Exact queries by name, symbol, filename or error text take a fast path. With few matches the agent immediately gets text and location; with many, a map of files and passages arrives first and the agent picks what to open. Several conditions go in `anyOf` and `allOf` lists, and the scope narrows to a folder. The agent continues a result and reads the context around a match by cursor: the next page, lines before and after the match. Search by meaning, the `concept` and `hybrid` modes, is off by default: it needs a local model and is enabled with `vectorSearchEnabled` in `sift-light.json`. The plugin works in Claude Code, Codex, Kimi Code, DeepSeek Harness, Pi and OMP, and connects to other clients as a regular MCP server. Built-in hooks by default forbid the agent ordinary shell search so that it uses the tool.
Who it is for. Developers whose agent burns context on repeated `grep` and file opens, and people who search large folders of notes, documents and logs.
Good fit when
- The agent keeps opening whole files to find a single line
- In a folder of contracts or notes you need to know which documents mention refunds
- A log search for an error message returns too many matches for one reply
- You remember the wording only roughly, and on-device search by meaning solves it
- You want one search rule across Claude Code and Codex as you work
Not a fit when
- The agent needs freedom to run any search command: in `hard` mode the hooks block ordinary shell search commands
- You need web search or cloud document search: the plugin searches the local disk only
- Node.js 22.19 or newer is missing: the server runs on it
- The machine is weak and you need semantic search all the time: the first model load takes tens of seconds and over 1 GiB of memory
Example request
Find mentions of refunds across all documents in the contracts folder, show the file map first and then open the two best matchesLimitations
The server requires Node.js 22.19 or newer. In the default mode hooks strictly forbid ordinary shell search; soften this with `SIFT_LIGHT_ENFORCE_SEARCH=prefer` or `off`. Semantic search is off, and once enabled the first local model load takes tens of seconds and over 1 GiB of memory; the authors call this an observation, not a guarantee. The license is AGPL-3.0. The adapters for Kimi Code and DeepSeek Harness 0.2.1-rc.5 are tied to specific client versions.
How to disable. In Claude Code disable the plugin through `/plugin`, in Codex through `/hooks` and the plugin manager, and for MCP remove the server with `claude mcp remove sift-light`. Hooks alone are switched off with `SIFT_LIGHT_ENFORCE_SEARCH=off`.
FAQ
How do I connect sift-light to Claude Code?
The easiest way is the plugin: `/plugin marketplace add lightsifter/sift-light`, then `/plugin install sift-light@sift-light`. For a bare MCP server run `claude mcp add sift-light -- npx -y --package sift-light@latest sift-light-mcp --stdio`. Restart the session after installing.
Why does the agent no longer run grep?
In the default `hard` mode the plugin hooks block ordinary shell search commands and steer the agent to `sift_light`. To keep both ways start the client with `SIFT_LIGHT_ENFORCE_SEARCH=prefer`; to lift the ban entirely set `off`.
How do I turn on search by meaning?
Install the local model with `npx -y --package sift-light@latest sift-light-model --install-model`, add `"vectorSearchEnabled": true` to `sift-light.json` and restart the client. Installing the model alone is not enough. Until then `concept` and `hybrid` return an explicit error and are never silently replaced by exact search.
What if a reply comes back with status waiting?
That is a long semantic search. Copy the `nextRequest` from the reply unchanged, usually `{"mode": "await", "operationId": "..."}`: it resumes the same work instead of starting over.
How do I connect sift-light to Codex?
Run `codex plugin marketplace add lightsifter/sift-light`, then `codex plugin add sift-light@sift-light`. After installing, review and confirm trust for the hook in `/hooks`. The Codex plugin includes a `local-search` skill that helps the agent find the deferred tool.
Discussion
No questions or reviews yet.
Related
PluginsOry Lumen
Lumen
A semantic code search plugin for Claude Code, Codex, Cursor and OpenCode that runs entirely on your machine
A local MCP server builds a SQLite and tree-sitter code index so an agent finds symbols, calls and meaning in one query
Local code search for Claude Code and Codex: six fast tools in place of grep and Read, with an index that updates itself.
codesearch searches code across several repositories with hybrid retrieval and runs fully offline, without a GPU or Docker.
Local code index for agents: the MCP tool map searches by meaning and structure, and code never leaves your machine
Source search ranks matches using code structure and gives agents relevant snippets without a separate search index.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| sift-light (this card) | no | 103 | ||
| grepai | no | 1,950 | searches code semantically and builds call graphs, without hooks that replace ordinary search | |
| Claude Context | yes | API key | 12,689 | indexes a codebase in a vector database rather than searching any local files by an exact path |
README in short
The README opens with the image of an attentive librarian and walks through scenarios: find a passage, start with a map, continue by cursor, find by inexact wording, search by several conditions and in a chosen folder. Then come install steps for Pi, OMP, MCP, Claude Code, Codex, Kimi and DeepSeek Harness, the search policy setup and local model parameters.