Documentation

Model Mapping

Which model category — Coding or General Purpose — each feature is restricted to, and why.

LoxiaFlow doesn't let you point every feature at every model. Each feature is locked to one of the two categories described in AI Config Categories — Coding (CLI tools and the LoxiaFlow Agent loop, with real file/shell access) or General Purpose (API-based, read-only) — so a feature that touches your codebase can never accidentally run on a model that can't actually edit files. This mapping is enforced in Settings → AI Config → Model Mapping, not just documented here — a model that isn't in the right category won't show up as a selectable option for that feature.

The quality of the output entirely depends on the model you use. For best results, use Claude — CLI or API — over the other options below.
FeatureModel categoryTier
01Planning & Requirements
DocumentsGeneral PurposeLight
02System Design
Style GuideGeneral PurposeLight
MockupsGeneral PurposeHeavy Generation
Project Layers — Run SetupCodingHeavy Coding
03Implementation
Generate WorkflowsGeneral PurposeHeavy Generation
Workflows / Workflow Runner (Run Layers)CodingHeavy Coding
04Testing & Security
Test GeneratorGeneral PurposeMedium Coding
Security ScansGeneral PurposeLight
05Maintenance
IssuesCodingHeavy Coding
Issue RunnerCodingHeavy Coding
Cross-Feature Utilities
AI ConsoleGeneral PurposeLight
Git ChangesNot AI-driven
TerminalNot AI-driven

Suggested models by tier

Every feature above maps to one of these four tiers — the list below is the suggested model/mode combination for that tier, shown once rather than repeated per feature.

Light
Haiku (Anthropic API)preferredHaiku (Claude CLI)Flash Medium (API)Any OpenAI-compatible provider (API)Ollama (Local)Ollama (Cloud)
Medium Coding
Haiku (Anthropic API)preferredHaiku (Claude CLI)Flash Medium (API)Any OpenAI-compatible provider (API)Ollama (Local)Ollama (Cloud)
Heavy Generation
Sonnet (Anthropic API)preferredSonnet (Claude CLI)Flash High (API)Any OpenAI-compatible provider (API)Ollama (Local)Ollama (Cloud)
Heavy Coding
Sonnet Code (Claude CLI)preferredFlash High (CLI)Ollama (Local · Agent loop)Ollama (Cloud · Agent loop)

Why the split

Project Layers (Run Setup), Workflows, Workflow Runner, Issues, and Issue Runner all drive the LoxiaFlow Agent loop against your real project files — so only agent-capable, Coding-purpose models (CLI tools like Claude CLI or Agy CLI, or an Ollama model with the Agent loop enabled) are selectable there. Every other AI feature runs a single prompt/response with no file or shell access — Style Guide, Documents, Mockups, Generate Workflows, AI Console, Test Generator, and Security Scans — so they're restricted to General Purpose models instead, including the plain Anthropic API and any OpenAI-compatible provider.

Git Changes and Terminal don't call a model at all — they're plain, non-AI utilities used alongside the AI-driven features above.

The model names in “Suggested models” mirror the default hints shown next to each feature in Settings → AI Config → Model Mapping. The (mode) tags are this page's own addition on top of that: General Purpose features can run the same model via its direct API (marked preferred — the app itself notes a direct API call is faster than going through a CLI) or via Claude CLI in restricted, tool-free mode, plus Ollama Local or Cloud. Coding features skip the API option entirely — a plain API/Anthropic config is always General-purpose and never agent-capable, so only CLI tools and an Ollama model with the Agent loop enabled are selectable there; Claude CLI is marked preferred there as the most capable, best-tested coding engine. These are starting points, not requirements — any configured model in the right category and mode works.