Skip to content

configuration.md

Latest commit

 

History

History
133 lines (96 loc) · 4.78 KB

configuration.md

File metadata and controls

133 lines (96 loc) · 4.78 KB
Raw

Configuration

AutoGPT is configured entirely through environment variables, no config files required. The table below lists every variable with its purpose, whether it is required, and which feature flag it belongs to.

Core Variables

Variable Required Default Description
AI_PROVIDER No gemini Active LLM provider: gemini, openai, anthropic, xai, cohere, huggingface
AUTOGPT_WORKSPACE No . (cwd) Workspace root where agents write generated files
ORCHESTRATOR_ADDRESS CLI only 0.0.0.0:8443 TCP address the orchestrator listens on

LLM Provider API Keys

Configure the key for your chosen provider. Only the key matching your AI_PROVIDER value is required.

# Gemini (default), requires feature: gem
export AI_PROVIDER=gemini
export GEMINI_API_KEY=<your_gemini_api_key>

# Override the Gemini model (otherwise uses first in provider list)
export GEMINI_MODEL=gemini-2.5-pro-preview-05-06

# OpenAI, requires feature: oai
export AI_PROVIDER=openai
export OPENAI_API_KEY=<your_openai_api_key>
export OPENAI_MODEL=gpt-4o   # optional override

# Anthropic Claude, requires feature: cld
export AI_PROVIDER=anthropic
export ANTHROPIC_API_KEY=<your_anthropic_api_key>
export ANTHROPIC_MODEL=claude-opus-4-6   # optional override

# XAI Grok, requires feature: xai
export AI_PROVIDER=xai
export XAI_API_KEY=<your_xai_api_key>
export XAI_MODEL=grok-4   # optional override

# Cohere, requires feature: co
export AI_PROVIDER=cohere
export COHERE_API_KEY=<your_cohere_api_key>
export COHERE_MODEL=command-r-plus   # optional override

# HuggingFace, requires feature: hf
export AI_PROVIDER=huggingface
export HF_API_KEY=<your_hf_api_key>
export HF_MODEL=meta-llama/Llama-3.3-70B-Instruct   # optional override

Obtain a Gemini API key from Google AI Studio. Keys for other providers are available on their respective developer portals.

Workspace

GenericGPT defaults to the current directory where the CLI is launched as its workspace. Set AUTOGPT_WORKSPACE to override:

export AUTOGPT_WORKSPACE=/my/project
autogpt          # file operations are scoped to /my/project

For the classic multi-agent workflow (BackendGPT, FrontendGPT, etc.), all agents write to subdirectories:

workspace/
├── architect/   # ArchitectGPT: diagram.py and generated PNGs
├── backend/     # BackendGPT: main.py, template.py, etc.
├── frontend/    # FrontendGPT: HTML/CSS/JS files
└── designer/    # DesignerGPT: image assets

The .autogpt Directory

GenericGPT maintains persistent state in a hidden directory inside the active workspace:

.autogpt/
├── sessions/          # Markdown conversation snapshots, auto-saved after every response
│   ├── <uuid>.md
│   └── ...
└── skills/                 # Learned lessons, injected on future sessions (TOML)
    ├── rust.toml
    ├── web.toml
    └── python.toml

Lesson files are plain TOML and can be viewed or manually edited. See the GenericGPT agent documentation for details.

Orchestrator Address

When running in Orchestrated Mode, the orchestrator (orchgpt) needs a bind address and the agent (autogpt --net) needs to know where to connect:

# On the orchestrator host
export ORCHESTRATOR_ADDRESS=0.0.0.0:8443

# On the agent host (must point to the orchestrator)
export ORCHESTRATOR_ADDRESS=192.168.1.10:8443

Optional: DesignerGPT (img)

Enables AI-generated image creation via GetImg:

export GETIMG_API_KEY=<your_getimg_api_key>

Optional: MailerGPT (mail)

Enables email reading and sending via Nylas:

export NYLAS_SYSTEM_TOKEN=<your_nylas_system_token>
export NYLAS_CLIENT_ID=<your_nylas_client_id>
export NYLAS_CLIENT_SECRET=<your_nylas_client_secret>

Optional: Long-Term Memory (mem)

Persists agent memory in Pinecone vector database:

export PINECONE_API_KEY=<your_pinecone_api_key>
export PINECONE_INDEX_URL=<your_pinecone_index_url>

The index URL looks like: https://my-index-abcdef.svc.us-east-1-aws.pinecone.io

💡 Tip Store secrets in a .env file and load them with source .env or a tool like direnv. Never commit API keys to version control.