Multi-Model Ensemble

NOME's ensemble includes Anthropic Claude (Opus/Sonnet), OpenAI (GPT-5.4/o-series), DeepSeek (Coder/V3), Google Gemini (Gemini/Vertex AI), and MiniMax (Conversational).

BYOK supports any OpenAI-compatible endpoint, allowing custom or self-hosted models to participate in the ensemble.

The universal router classifies each request by workload type, complexity, and required capabilities. Simple queries route to fast, cost-efficient models. Complex architecture and analysis tasks route to primary reasoning engines.

Routing is deterministic-first with LLM fallback — no public framework matches this approach for reliability.

Each HarnessRunFrame carries explicit effort budgets: max subagents, max tool calls, and max token budget. These prevent over-investment in simple queries.

Ensemble aggregation supports multiple strategies: best_of_n, consensus_vote, merge_and_refine, and first_success.

Connect any OpenAI-compatible API endpoint as a custom model provider. NOME routes to it using the same harness contract, receipts, and cost tracking as built-in providers.

Custom endpoints participate in failover chains and can be targeted by routing policies for specific workload types.

NOME tracks input and output tokens across all connected models simultaneously. Metrics are normalized into clear visual dashboards with per-provider and per-run breakdowns.

The cost audit surface shows real-time spending, historical usage trends, and per-surface token consumption.

When a provider approaches its rate limit or returns errors, NOME automatically transitions to a redundant provider based on the configured fallback chain.

Failover is transparent to the user and produces audit events. The harness contract is preserved across provider switches — receipts still track which provider handled each segment.