OmniSeek

What the OmniSeek MCP server does

The OmniSeek MCP server gives an MCP-connected model access to a self-hosted perception and research system. Its catalog supports search across configured sources, URL and document extraction, image and video-frame interpretation, and local audio or video transcription. It also includes tools for scholarly discovery, identity resolution, collaboration analysis, institution cohorts, paper enrichment, and semantic recommendations.

The system is intended for investigations where ordinary keyword search is insufficient. Examples include finding information in another language, extracting text from PDFs, locating evidence in audio or video, examining research fields, and tracking changes in a set of sources. A persistent retrieval memory and typed evidence graph allow information gathered over time to be queried and related.

How it works

An MCP client connects to the running service, while the model and the client-side agent loop decide which tools to call. omniseek_search fans out across the available catalog, removes duplicates, and ranks results using lexical and semantic methods. omniseek_sources reports available domains, regions, capabilities, and source health so an agent can select an appropriate route.

Content can then pass through omniseek_read, omniseek_view, or omniseek_transcribe. Reading handles web pages, PDFs, arXiv material, and other documents; viewing handles images, document figures, and video frames; transcription runs locally and supports bilingual speech plus timestamp ranges. Cross-lingual semantic retrieval depends on the index built as the installation is used, so a new installation begins with less accumulated recall.

Research tools build on this content. Graph queries can find entities, neighborhoods, paths between items, time-bounded results, and similar items. Other tools record identity decisions and directed relations, while sensors issue standing queries and return only newly detected material. Source curator tools manage submission, probing, judgment, admission, retirement, and audit views.

Setup and configuration

Docker is the recommended installation path. The compose setup starts an HTTP service on the local machine, exposes a health endpoint, and prints a bearer token on first startup. The MCP endpoint is available at http://127.0.0.1:8765/mcp; credentials are stored under the OmniSeek credentials directory. The service binds to localhost and expects the bearer token on every request.

A Python virtual environment can run the service without Docker. The base installation provides core keyless APIs, static sources, document reading without PDF support, and lexical memory. Optional extras add PDF support, ASR, cross-lingual vectors, OCR, and login-walled sources. Login-walled access requires accounts supplied by the operator. OMNISEEK_CONTACT_EMAIL can be set to request a faster lane with Crossref, SEC, and Unpaywall.

The package also provides an omniseek stdio command for clients that launch MCP servers themselves. Windows users should use Git Bash or WSL for the bootstrap script, or use Docker instead.

Tools and capabilities

The OmniSeek MCP server exposes tools for:

• Catalog search, source routing, parallel calls, and document normalization.
• Vision-based reading and local bilingual audio/video transcription.
• Citation-field mapping, author identity candidates, coauthor networks, institution cohorts, paper metadata, and SPECTER-based recommendations.
• Evidence-graph queries, novelty monitoring, identity rulings, and carried-forward statements.
• Source admission workflows and per-source audit dossiers.

Limitations and notes

Optional capabilities are not present in every installation. The prebuilt core image omits PDF reading, ASR and video-frame hearing, and login-walled sources; those require a local build or additional installation choices. Login-walled sources remain disabled until the operator supplies accounts. Do not expose the local service directly without a reverse proxy. The README does not identify specific MCP client products, so client compatibility should be verified against the chosen client’s stdio or HTTP support.