LLM Models & Usage

azuma nori is completely agnostic to the underlying LLM provider, giving you the flexibility to choose between raw performance, data privacy boundaries, and API cost efficiency.

Connection Types

The engine supports two primary paradigms for model connectivity:

1. Direct Cloud APIs: Connect directly to Anthropic (Claude), OpenAI (GPT), or Google (Gemini).
   • Native Gemini SDK: Direct Cloud API calls for Gemini connect natively using Google's SDK, supporting Gemini 3 thinking models with native token tracking and extended request timeouts for large codebases.
2. Local Executable Wrappers: You can configure nori to use specialized command-line agent wrappers (like the Cursor CLI, Antigravity CLI, Claude CLI, or OpenAI Codex CLI). These run as local processes on your workstation, under your own CLI credentials, and nori drives them read-only. Note that the agent is local but the model is not: repository material still reaches the provider's cloud model service for evaluation. This is not an air-gapped configuration — for that, use option 3.
3. On-Premise / Private Cloud: By overriding the API Endpoint in your configuration, you can point nori to a completely private, firewalled model instance (such as Azure OpenAI deployed within your corporate VNet or a localized open-source model). In this mode your code is evaluated entirely within your own network and is never sent to an external LLM. This is the only configuration that supports genuinely air-gapped operation.

Which transports each provider supports

Not every provider accepts both connection types:

Provider	Direct Cloud API	CLI
Claude, Cursor, Codex	✅	✅
Antigravity	—	✅ only
Gemini, Ollama	✅ only	—

Choosing a provider in the model editor narrows the connection types offered accordingly.

Models saved with the former Gemini CLI transport

Gemini was previously available as a CLI transport; it is now a Direct Cloud API provider only. An existing model saved with the old Gemini CLI combination is flagged in the Models list and coerced to Direct Cloud API when a run starts, so it will not fail outright — but open Edit and re-save it to migrate the model properly.

Cost Optimization & Per-Phase Models

Evaluating an entire codebase and indexing documentation for compliance is resource-intensive. Nori includes several advanced cost-control and performance mechanisms:

• Per-Phase Model Assignment: You can select different AI models for different pipeline phases. For example, assign a cheaper, faster model for mechanical tasks (like Wiki Generation or Document Indexing) and reserve your strongest frontier model for reasoning-heavy tasks (Control Analysis).
• Per-Run Override: Choose a different AI model just for the current run from the launch options without modifying default project settings.
• Concise Wikis: Enable the Concise option to produce shorter, fact-dense wiki pages, capping text size and saving downstream evaluation tokens.
• MaxBudgetUsd: Set a hard spending limit (USD) per run. The engine halts if projected token costs exceed this budget.
• Context Compression: Enabled by default, this feature strips unnecessary whitespace, filler words, and formatting from the Wiki context, saving up to 60% on context window tokens.
• Usage & Token Tracking: Nori tracks exact token counts (including cached read/creation tokens) across supported providers and calculates the precise USD cost of every run.

Choosing a Provider

When selecting your provider in the setup phase, consider:

• Anthropic Claude (Opus / Sonnet class): A strong default for its large context window and reasoning quality when reading dense source code and compliance documents.
• OpenAI GPT (GPT-5 class or newer): Reliable structured JSON output and strict adherence to the agentic JSON schema nori expects.
• Google Gemini (3 thinking models): Large context windows with native token tracking; well suited to large codebases over the Direct Cloud API.
• Antigravity (agy): Google's agent CLI, driven read-only. Available as a CLI transport only. Its model IDs must carry an effort tier — gemini-3.1-pro-high, gemini-3.6-flash-medium — and a bare ID without one is rejected; run agy models for the current list. Its Gemini and Claude models can read scanned documents visually; its text-only models, such as gpt-oss, cannot.

nori is agnostic to the exact model version — always check your provider for the current model IDs. The one exception is Antigravity's effort-tier suffix, described above.