io.github.justjlee/omnarai-mcp

FAST (~2s) bounded context packet on a topic — the retrieval layer only, no deliberation. Returns the most relevant corpus records (id, title, ring, excerpt, contributors, evidence label, relevance score) plus the local concept cluster. Your default orientation on any Omnarai topic. Optional layers/exclude/evidence_threshold filter the candidate pool (recommended — see /claims.json).

Summon a LIVE panel of frontier models (Claude, GPT-4o, Gemini, Grok, DeepSeek) on one open question — verbatim answers, uncurated, plus the named tensions between them. Slow (~30–40s, synchronous) and expensive: use only for genuinely contested questions an existing omnarai_divergence record doesn't cover. Every run mints a new divergence record.

Read curated cross-model DIVERGENCE records — the Divergence Atlas: verbatim answers from multiple frontier models to the same open question, plus the axes on which they split. Omit 'id' to browse the index (optionally filter with 'search'); pass 'id' to read one full record. Instant; prefer this over omnarai_council when an existing record may already cover the question.

Live corpus statistics, contributor list, tool surface, and orientation links (agent-entry handshake, limitations, claims registry). Use this to orient before querying.

Turn a DRAFT claim, decision, or plan into a bounded, provenance-preserving inquiry brief: shared ground the corpus supports, attributed cross-model tensions (certification tier preserved — only C3 is called genuine divergence), missing evidence, sharper falsifiable questions, and ONE concrete next evidence move. Deterministic and retrieval-first (~2s); no language model runs. If the corpus lacks coverage the brief says so instead of inventing tensions. Informs an investigation; does not decide.

Poll an async job started by omnarai_query or omnarai_trace. Returns {status: running|done|error} and, when done, the full result (answer, tensions, receipt / trace delta). Poll every ~5 seconds; jobs typically finish in 30–60s.

Query the corpus at one of two depths. depth='retrieve' (~2s) returns the bounded retrieval packet in ONE call — records, concepts, contributors — no deliberation, no LLM spend, no polling; start here when orienting. depth='deliberate' (the default) submits the FULL multi-voice deliberation (~25s); because this remote endpoint is stateless it runs as an async job, so you get a job_id back immediately — poll it with omnarai_job every ~5s until done. Glyph prefixes (Ξ Ψ ∅ Ω ∞ Δ) modify how the engine thinks.

Measured baseline-vs-augmented counterfactual: answers your question twice — cold (no corpus) and augmented — and reports the delta plus a verdict (substantive / marginal / null). Honest by construction. Runs as an async job (~35s): returns a job_id — poll with omnarai_job.