im-looking-for for AI agents

English · 한국어

im-looking-for lets an agent find a file on the user's computer by what's in it — "the photo of the whiteboard from the offsite", "the PDF where the lease talks about pets", "the voice memo about the dentist" — across photos, scans, slides, documents, audio and video, in 100+ languages. Everything runs locally.

There are two ways in: the MCP server (ilf mcp) and the CLI with --json. Both report the same JSON shapes.

Connect

If ilf isn't on the agent's PATH, use the absolute path. The app's Library panel shows the exact command for this install, with a Copy button.

The user has to run setup once (ilf, or ilf setup) and add folders before there is anything to search. An agent can add folders and start indexing itself (add_folder, index), but the ~900 MB model download is left to the user.

MCP tools

index returns at once. Poll status until indexing is false; progress has total, read, unchanged, unreadable and the current file. Searching works during indexing, over what's been read so far.

Read-only tools carry readOnlyHint; remove_folder carries destructiveHint (it only edits the folder list, never files on disk).

Result

{
  "path": "/home/you/Documents/lease_2026.pdf",
  "name": "lease_2026.pdf",
  "kind": "pdf",
  "score": 0.7765,
  "match": "page",
  "loc": 1.0,
  "loc_end": null,
  "where": "p.1",
  "snippet": "주택 임대차 계약서 제 1 조 (목적) …"
}

One result per file, best match first.

Reading scores

A search always returns the nearest files, even when nothing matches. Measured on the demo folder: lease termination → the lease PDF at 0.78, then unrelated files at 0.59; the nonsense query xyzzy-nothing-here still returns a video at 0.61. So:

• A score means something only next to the others. A result well ahead of the rest (here 0.78 vs 0.59) is a match; a flat run of similar scores means "probably not here".
• Check before you claim: read the snippet, or call read_passages with the result's path and loc.
• Never tell the user a file doesn't exist because search didn't find it. Its folder may not be added or indexed (status), it may be a format that isn't read (HEIC, code), or the words may be inside a photo (no OCR).

Writing queries

• Describe content, not names: a cat sleeping on a red sofa, not IMG_2041. The file name is searched too, as a keyword, so a known name still helps.
• Any language works, and languages mix: an English query finds a Korean PDF.
• Use kinds when the user said what it is ("the recording…", "that PDF…").
• Exact words count: a word that appears literally in a document lifts it, which helps with names, codes and numbers.

CLI with --json

ilf invoice from march --json            # same as: ilf search invoice from march --json
ilf -k image -n 5 whiteboard --json
ilf status --json
ilf add ~/Scans --json && ilf index --json

ilf index --json blocks until done and prints {read, unchanged, forgotten, unreadable, seconds, errors}. Progress goes to stderr only on a terminal, so stdout stays clean JSON.

Status

{
  "version": "0.1.0",
  "ready": true,
  "data_dir": "/home/you/.local/share/im-looking-for",
  "runtime": "llama-b11461-bin-ubuntu-x64.tar.gz",
  "backend": "cpu",
  "gpu_failed": null,
  "backends": ["cpu", "vulkan"],
  "missing_models": [],
  "download_bytes": 0,
  "folders": ["/home/you/Pictures", "/home/you/Documents"],
  "excluded_folders": [],
  "excluded_extensions": [],
  "index": {"files": {"image": 6, "pdf": 1, "doc": 3}, "errors": 0, "vectors": 17, "bytes": 98304}
}

Privacy

Nothing leaves the computer: the model runs in a local llama.cpp process and the index is a local SQLite file. read_passages returns only text from files already in the index, which the user chose by adding their folders.