AI Backends

Contents

• ClaudeCode (Recommended)
  • Setup
  • Configure as default
• OpenAI-compatible
  • Setup
  • Azure AI Foundry
• Ollama
  • Setup
  • Model selection notes
• Changing Backends Mid-project

OpenTransmute supports three AI backends for decompose, compose, and implement jobs. All three implement the same internal interface — switching backends requires only a settings change.

---

ClaudeCode (Recommended)

Invokes the claude CLI as a managed subprocess. The agent uses its own file-reading tools to explore the source directory during decomposition — no file packing or chunking required. Model weight tiers are automatically mapped to Anthropic’s model families.

Weight tier	Maps to
Thick	Claude Opus
Regular	Claude Sonnet
Thin	Claude Haiku

Setup

# Install the Claude Code CLI
npm install -g @anthropic-ai/claude-code

# Authenticate (follow the prompts)
claude login

# Verify
claude --version

Configure as default

otx settings --orchestrator ClaudeCode

No API key is required — authentication is handled by the claude CLI itself. Model names cannot be overridden for this backend; Anthropic’s tier mapping is used automatically.

---

OpenAI-compatible

Calls any OpenAI-compatible HTTP API. You supply the base URL and API key; OpenTransmute handles the rest.

Compatible providers:

Provider	Base URL
OpenAI	https://api.openai.com/v1
Azure AI Foundry	https://<resource>.openai.azure.com/openai/deployments/<deployment>
LM Studio (local)	http://localhost:1234/v1
vLLM	http://localhost:8000/v1
Any OpenAI-compatible proxy	Your endpoint URL

Setup

otx settings \
  --orchestrator OpenAI \
  --endpoint https://api.openai.com/v1 \
  --thick-model gpt-4o \
  --regular-model gpt-4o-mini \
  --thin-model gpt-4o-mini

Pass the API key at runtime (never saved to disk):

otx decompose https://github.com/org/repo --api-key sk-...
# or
export OPENAI_API_KEY=sk-...
otx decompose https://github.com/org/repo

Azure AI Foundry

For Azure, include the deployment in the base URL and use the Azure API key:

otx settings \
  --orchestrator OpenAI \
  --endpoint "https://my-resource.openai.azure.com/openai/deployments/my-deployment"
otx decompose ./my-project --api-key <azure-api-key>

---

Ollama

Calls a local Ollama instance via its OpenAI-compatible API at http://localhost:11434. No API key required — full privacy, data never leaves your machine.

Setup

1. Install Ollama for your platform.
2. Pull models for each weight tier:

# Example: using Llama 3.1 variants
ollama pull llama3.1:70b   # thick
ollama pull llama3.1:8b    # regular
ollama pull llama3.2:3b    # thin

1. Configure:

otx settings \
  --orchestrator Ollama \
  --thick-model llama3.1:70b \
  --regular-model llama3.1:8b \
  --thin-model llama3.2:3b

1. Confirm Ollama is running before starting a job:

ollama list

Model selection notes

• Phase 3 (Component Specifications) is the most demanding phase — use the largest available model for the thick tier.
• For the thin tier (Phase 0, Phase 5, Phase 6, Phase 7), a 3B–8B model is usually sufficient.
• Context window size matters significantly for large codebases. Prefer models with at least 32k token context for the thick tier.

---

Changing Backends Mid-project

You can change backends between phases without losing work. The decompose job state and all completed phase outputs are saved to disk. To resume with a different backend:

otx decompose https://github.com/org/repo \
  --project my-project \
  --start-phase 4 \
  --orchestrator ClaudeCode