Omneum

What Omneum MCP server does

Omneum MCP server adds a post-retrieval evaluation step to an agent pipeline. It analyzes relationships between observations, claims, and their supporting sources, then produces reliability and support scores. The analysis can account for source dependencies instead of treating repeated information as independent evidence. This helps an agent distinguish genuinely corroborated context from several records that ultimately came from the same upstream material.

The evaluator does not interpret the semantic meaning of a claim. Instead, scores are derived from the structure and metadata associated with observations, including provenance and lineage. Omneum uses the BRAID algorithm to calculate the resulting signals. The agent remains responsible for deciding how to use those signals in its workflow.

How it works

A typical pipeline places the evaluator after retrieval and optional reranking:

• Retrieve records or context.
• Optionally rerank the retrieved material.
• Map records into Omneum observations.
• Send the mapped observations to the local evaluator through MCP.
• Use the returned quality signals in the agent’s control flow.

The Python SDK provides ContextMapper for converting application records into observations. A mapping can define source identity, entity, attribute, value, observation ID, observation time, and upstream sources. The SDK also handles canonicalization. Dependency estimation can use factors such as upstream relationships, citations, assertion lineage, ownership, temporal relationships, graph structure, and retrieval metadata.

Omneum MCP server runs locally and does not add an inference request or another model call to the retrieval path. Its evaluator runs as a separate system service, while the MCP process connects through omneum-server.

Setup and configuration

Install the Python package with pip, then run the installer from a normal login account:

pip install omneum
omneum install

The documented quick start supports Python 3.11 through 3.14, macOS on Apple Silicon or Intel, and Ubuntu 22.04 or newer on x86_64. The installer downloads and verifies a version-pinned evaluator release. On supported systems, the evaluator is installed as a managed service: launchd on macOS or systemd on Ubuntu. Administrator access may be requested for machine-level service files.

Add the stdio server to an MCP client configuration:

{
  "mcpServers": {
    "omneum": {
      "command": "omneum-server"
    }
  }
}

Restart the client and call the ping tool to confirm connectivity. If the executable is not found, configure the absolute path returned by command -v omneum-server.

Tools and capabilities

The documented MCP tools and SDK operations include:

• ping, which verifies that the client can reach the local service.
• evaluate_mapped(), which evaluates records after they have been mapped into observations.
• ContextMapper, which defines how application records become Omneum observations.
• Dependency estimator configuration for weighting provenance, lineage, temporal, graph, and related signals.
• Reliability and claim-support scoring based on source relationships and available metadata.

Limitations and notes

Omneum MCP server currently documents Python SDK support rather than SDKs for other languages. The quick-start installation is limited to the listed macOS and Ubuntu environments. The evaluator depends on provenance and lineage metadata supplied with retrieved records; incomplete metadata can limit dependency estimation.

The system evaluates structural support and reliability signals without understanding the semantic truth of the text itself. It therefore complements retrieval and ranking rather than replacing them. A running service is not sufficient to verify installation: the installer also checks a request through the MCP path and validates the response against the configured public key.