@hunterstrategy/sift

Token-efficient, structured retrieval over code, docs, and tabular data, exposed to AI agents via MCP. Instead of reading whole files, sift parses sources into addressable units — sections, symbols, rows — and serves back exactly the one unit requested. search returns lightweight headers and IDs; retrieve fetches the byte-exact unit, hash-verified. Models spend context on content, not navigation.

Install

Requires Node.js ≥ 22.13.

npm i -g @hunterstrategy/sift
sift-init

sift-init writes a single global sift server entry into your AI client's MCP config (Claude Code, Claude Desktop, Cursor, …). Run it once, from anywhere. Your client then launches the server itself, and a project indexes itself the first time you reference it — no per-project setup.

sift keeps one personal index at ~/.sift that spans every project you've referenced.

Binaries

| Binary | Purpose | |---|---| | sift-all | MCP server — all three domains + cross-domain join | | sift-docs | MCP server — docs domain only | | sift-data | MCP server — data domain only | | sift-code | MCP server — code domain only | | sift-init | Write or merge sift into your AI client's MCP config | | sift-index | Index (or refresh) a project directory on demand | | sift-watch | Watch a root directory and re-index on change | | sift-doctor | Report index health: drift, missing files, schema version, dead roots. --reindex clears stale indexes after a schema bump; --prune removes indexes for projects that no longer exist |

Server binaries accept an optional path argument (sift-all /path/to/project); omitting it defaults to the current directory.

Supported formats

| Domain | Formats | |---|---| | Docs | Markdown, HTML, MDX, RST, AsciiDoc, OpenAPI, JSON, XML, Jupyter Notebooks, PDF, Word (.docx) | | Data | CSV, TSV, Excel (.xlsx), Parquet, JSONL/NDJSON | | Code | TypeScript, JavaScript, Python, Go, Java, Rust, C#, Bash, HCL/Terraform |

For PDF and DOCX, retrieve returns extracted text rather than a raw byte slice; those units are marked contentMode: "derived-text". A scanned, image-only PDF yields a single unit flagged extractable: false rather than silent empty content.

MCP tools

discover, search, retrieve, describe, rows, join, semantic-search, index.

License

MIT © Hunter Strategy LLC