Skip to content

feature-flags.md

Latest commit

 

History

History
84 lines (61 loc) · 4.15 KB

feature-flags.md

File metadata and controls

84 lines (61 loc) · 4.15 KB
Raw

Feature Flags

AutoGPT's Cargo.toml uses feature flags to keep the binary lean. Only compile what you need. Features are additive: enable multiple with a comma-separated list.

Full Feature Table

Feature Enables Required Env Vars
gem Gemini LLM client (gems crate) GEMINI_API_KEY
oai OpenAI client (openai_dive crate) OPENAI_API_KEY
cld Anthropic Claude client ANTHROPIC_API_KEY
xai XAI Grok client (x-ai crate) XAI_API_KEY
co Cohere client (cohere-rust crate) COHERE_API_KEY
hf HuggingFace client (api_huggingface crate) HF_API_KEY
mcp Model Context Protocol (MCP) support N/A
gpt All built-in GPT agents (ManagerGPT, ArchitectGPT, BackendGPT, FrontendGPT, OptimizerGPT) One LLM feature
img DesignerGPT image generation (getimg crate) GETIMG_API_KEY
git GitGPT auto-commit (git2 crate) Git repo in workspace
mail MailerGPT email integration (nylas crate) NYLAS_* vars
mem Pinecone long-term memory (pinecone-sdk crate) PINECONE_API_KEY, PINECONE_INDEX_URL
net IAC protocol networking (iac-rs crate) ORCHESTRATOR_ADDRESS
cli Interactive shell, session mgmt, YAML/TOML parsing, terminal UI One LLM feature

Common Combinations

Minimal SDK, Gemini only:

autogpt = { version = "0.4.5", features = ["gem", "gpt"] }

Full CLI with Gemini:

autogpt = { version = "0.4.5", features = ["gem", "gpt", "cli"] }

SDK with memory:

autogpt = { version = "0.4.5", features = ["gem", "gpt", "mem"] }

Networked orchestrated mode:

autogpt = { version = "0.4.5", features = ["gem", "gpt", "net", "cli"] }

Everything:

autogpt = { version = "0.4.5", features = ["gem", "gpt", "img", "git", "mail", "mem", "net", "cli"] }

Or equivalently from the command line:

cargo install autogpt --all-features

Compile-Time Feature Guards

In the source code, all feature-gated APIs are guarded with #[cfg(feature = "...")]. If you call DesignerGPT::new without the img feature, compilation fails with a clear #[cfg] error rather than a runtime panic.

The AI_PROVIDER Environment Variable

At runtime, AI_PROVIDER selects the active LLM client among all compiled-in providers:

export AI_PROVIDER=gemini     # uses gem feature
export AI_PROVIDER=openai     # uses oai feature
export AI_PROVIDER=anthropic  # uses cld feature
export AI_PROVIDER=xai        # uses xai feature
export AI_PROVIDER=cohere     # uses co feature
export AI_PROVIDER=huggingface # uses hf feature

If AI_PROVIDER is not set, AutoGPT defaults to gemini.

💡 Tip Enabling only the features you need significantly reduces compile time and final binary size. The release profile is already optimized for size (opt-level = "z", LTO, symbol stripping).