# OmniSeek

A self-hosted MCP server that gives an AI agent perception beyond text search: local speech transcription, reading behind a login wall with the user's own credentials, seeing images and video frames, crossing languages, and remembering past findings.

OmniSeek extends what an agent's search can reach past what a plain search index covers, in one language, in text: it transcribes audio locally with bilingual automatic speech recognition, reads images and video frames, translates across languages so a query in one language can surface results in another and back, and can read content behind a login using the user's own credentials on the user's own machine, off by default. Everything it finds is kept in a persistent retrieval memory alongside a typed, source-traced evidence graph, and cross-lingual results improve as the index it builds grows with use, so a fresh install represents the coldest, not the typical, case.

The project frames its value against a plain search tool that stops once it has quoted an official rule: OmniSeek can go on to surface first-person accounts sitting behind a login, or a spoken explanation captured only in a video with no separate transcript, naming the sources it held back with the exact call needed to retrieve each one.

It is written in Python and released under the Apache-2.0 license, with documentation in English, Chinese and Japanese.

Website: https://battam1111.github.io/omniseek/