OmniSeek

The answer is sitting in minute 47 of a podcast, three replies deep in a comment thread, behind a login, in another language. Your agent gets it anyway.

_{Self-hosted perception MCP server · one connection}

       

Quick start · Tools · Configure · Contributing

Languages: English · 中文 · 日本語

Search gives your agent indexed pages, in one language, in text, and stops there.

OmniSeek gives it the senses to keep going: through languages, logins, comment threads, audio, and pixels, all on your machine.

What each layer gave back, verbatim:

• Written down, and in reach. Headlines, official FAQ, top blogs, all one voice: "From 2026, F-1 admission is limited to a 4-year initial period; renewal in a third country remains possible." All quote the same rule. None of them have done it.
• Written down, but out of reach. Three first-person threads on 1point3acres, behind your own login: Bangkok, booked to passport in 25 days, interview to approval in 30 minutes; Milan, a month-long fight for a slot, visa issued for 5 years; Tokyo, "silky-smooth". Under the Milan post the author comes back in the comments: "Book any late slot first, then email the consulate to expedite. For one F-1 applicant it worked." One person's experience, not official guidance.
• Never written down. A Chinese explainer video on bilibili, transcribed locally: the "4-year cap" in the headlines is the initial period, extensions moved desks rather than vanishing. A rednote video note whose caption is four hashtags, frames and speech read locally: a 212(a)(6)(C) refusal abroad, a misrepresentation finding, can nearly close the F-1 road.

Plain search quoted the rule and stopped. The people who had lived it held the timelines, the workaround, and the risk. OmniSeek also named the sources it held back, each with the exact call to drill it.

It hears (local bilingual ASR, no cloud), sees (images and video frames, in-band), crosses languages (a Chinese query finds English results and vice versa), reads behind login walls (your credentials, your machine, off by default), and remembers (persistent retrieval memory plus a typed, source-traced evidence graph).

Crossing languages draws on the index OmniSeek builds as you use it, so a fresh install starts at a floor. The published claim-verification tests run on exactly that fresh install, which makes their cross-lingual number the coldest case rather than the typical one.

Every source in the catalog, the curated roster of everything OmniSeek can reach, earned its place by beating plain search at one of five jobs (structure, unwalling, transcription, recall, monitoring): citation graphs, regulatory filings, login-walled forums, Chinese-language video. And the catalog is built to grow: a curator pipeline probes, judges, and admits new sources, and retires the ones that decay.

Worked examples, real outputs · A full case study · Every claim above is a test (latest results) · Source health, updated weekly

Quick start

Docker (recommended)

git clone https://github.com/Battam1111/omniseek.git && cd omniseek
docker compose up -d
docker compose logs omniseek        # bearer token printed on first start
curl -s http://127.0.0.1:8765/healthz

On success, it returns:

{"ok":true}

Point your MCP client at http://127.0.0.1:8765/mcp with Authorization: Bearer <token>. The token is generated on first start and stored in ~/.omniseek/credentials/omniseek_http.json (with the compose file, that's ./.omniseek/credentials/omniseek_http.json on the host).

Two paths from here. The prebuilt core image (docker pull ghcr.io/battam1111/omniseek, amd64 + arm64) needs no build and carries every core sense; it is Apache-clean and ships without PDF reading, hearing (ASR + video frames), and login-walled sources. Wanting those extras is what triggers a local build: set EXTRAS="[pdf,asr,walled]" and run docker compose build, then up -d (the first build also fetches headless Chromium; later starts are instant). Optional but recommended: set OMNISEEK_CONTACT_EMAIL for a faster lane with Crossref, SEC, and Unpaywall.

Without Docker

python -m venv .venv && . .venv/bin/activate
scripts/bootstrap.sh
python -m omniseek.serve_http

The bare install is the Core tier: every keyless API and static source, document reading minus PDF, and the lexical memory index. pip install "omniseek[pdf,asr,recall,ocr]" wakes the Research tier (PDF, hearing, cross-lingual vectors, OCR); omniseek[walled] adds the login-walled tier, which stays off until you bring your own accounts; omniseek[all] takes everything. The server prints which senses are online, and which are dormant, at every boot.

On Windows, run bootstrap.sh under Git Bash or WSL; Docker is the simplest path. For an always-on Linux service, see deploy/omniseek.service.

Prefer stdio? The install also ships an omniseek command that speaks MCP over stdio, for clients that launch servers themselves; Dockerfile.stdio wraps the same thing in a container.

OmniSeek binds 127.0.0.1 and requires the bearer token on every request. Do not expose without a reverse proxy (SECURITY.md).

Tools

One MCP connection; no model, no agent loop inside. Your model thinks, your harness drives the loop, OmniSeek reaches. Start with omniseek_search; explore what's available with omniseek_sources.

The login-walled tier has no tool of its own: once you opt in per source, the same omniseek_search(..., sources=["xiaohongshu"], raw=True) runs through your own logged-in browser. See walled sources.

Full reference in tools.md · FAQ

Using Claude Code? skills/omniseek-investigate ships the investigation methodology (sweep, zoom, structure) as a ready-made skill.

Configure

OmniSeek is catalog-first: with no config, every benign source is on and login-walled sources are off. Tune in one file, ~/.omniseek/profile.json (example):

Read the full README on GitHub →