Skip to main content

LLM & Environment Configuration

FuzzForge AI relies on LiteLLM adapters embedded in the Google ADK runtime, so you can swap between providers without touching code. Configuration is driven by environment variables inside .fuzzforge/.env.

Minimal Setup​

LLM_PROVIDER=openai
LITELLM_MODEL=gpt-5-mini
OPENAI_API_KEY=sk-your-key

Set these values before launching fuzzforge ai agent or python -m fuzzforge_ai.

.env Template​

fuzzforge init creates .fuzzforge/.env.template alongside the real secrets file. Keep the template under version control so teammates can copy it to .fuzzforge/.env and fill in provider credentials locally. The template includes commented examples for Cognee, AgentOps, and alternative LLM providers—extend it with any project-specific overrides you expect collaborators to set.

Provider Examples​

OpenAI-compatible (Azure, etc.)

LLM_PROVIDER=azure_openai
LITELLM_MODEL=gpt-4o-mini
LLM_API_KEY=sk-your-azure-key
LLM_ENDPOINT=https://your-resource.openai.azure.com

Anthropic

LLM_PROVIDER=anthropic
LITELLM_MODEL=claude-3-haiku-20240307
ANTHROPIC_API_KEY=sk-your-key

Ollama (local models)

LLM_PROVIDER=ollama_chat
LITELLM_MODEL=codellama:latest
OLLAMA_API_BASE=http://localhost:11434

Run ollama pull codellama:latest ahead of time so the adapter can stream tokens immediately. Any Ollama-hosted model works; set LITELLM_MODEL to match the image tag.

Vertex AI

LLM_PROVIDER=vertex_ai
LITELLM_MODEL=gemini-1.5-pro
GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json

Additional LiteLLM Providers​

LiteLLM exposes dozens of adapters. Popular additions include:

  • LLM_PROVIDER=anthropic_messages for Claude 3.5.
  • LLM_PROVIDER=azure_openai for Azure-hosted GPT variants.
  • LLM_PROVIDER=groq for Groq LPU-backed models (GROQ_API_KEY required).
  • LLM_PROVIDER=ollama_chat for any local Ollama model.
  • LLM_PROVIDER=vertex_ai for Gemini.

Refer to the LiteLLM provider catalog when mapping environment variables; each adapter lists the exact keys the ADK runtime expects.

Session Persistence​

SESSION_PERSISTENCE=sqlite   # sqlite | inmemory
MEMORY_SERVICE=inmemory      # ADK memory backend

Set SESSION_PERSISTENCE=sqlite to preserve conversational history across restarts. For ephemeral sessions, switch to inmemory.

Knowledge Graph Settings​

To enable Cognee-backed graphs:

LLM_COGNEE_PROVIDER=openai
LLM_COGNEE_MODEL=gpt-5-mini
LLM_COGNEE_API_KEY=sk-your-key

If the Cognee variables are omitted, graph-specific tools remain available but return a friendly "not configured" response.

MCP / Backend Integration​

FUZZFORGE_MCP_URL=http://localhost:8010/mcp

The agent uses this endpoint to list, launch, and monitor Prefect workflows.

Tracing & Observability​

The executor ships with optional AgentOps tracing. Provide an API key to record conversations, tool calls, and workflow updates:

AGENTOPS_API_KEY=sk-your-agentops-key
AGENTOPS_ENVIRONMENT=local     # Optional tag for dashboards

Set FUZZFORGE_DEBUG=1 to surface verbose executor logging and enable additional stdout in the CLI. For HTTP deployments, combine that with:

LOG_LEVEL=DEBUG

The ADK runtime also honours GOOGLE_ADK_TRACE_DIR=/path/to/logs if you want JSONL traces without an external service.

Debugging Flags​

FUZZFORGE_DEBUG=1           # Enables verbose logging
LOG_LEVEL=DEBUG             # Applies to the A2A server and CLI

These flags surface additional insight when diagnosing routing or ingestion issues. Combine them with AgentOps tracing to get full timelines of tool usage.

  • Env bootstrap: ai/src/fuzzforge_ai/config_manager.py
  • LiteLLM glue: ai/src/fuzzforge_ai/agent.py
  • Cognee integration: ai/src/fuzzforge_ai/cognee_service.py