Getting started¶

Install¶

pip install arc-agentkit

Distribution name vs import name

On PyPI the distribution is arc-agentkit (the agentkit name on PyPI is owned by an unrelated project). The import name stays agentkit, so all documentation and examples use import agentkit as-is. Same pattern as pip install Django → import django.

Before the first release, install directly from Git:

pip install "arc-agentkit @ git+https://github.com/arc-labs-ai/agentkit"

The core has zero runtime dependencies. Provider integrations live behind opt-in extras — pull only what you need:

pip install "arc-agentkit[http]"            # bundled httpx-based LLM providers
pip install "arc-agentkit[postgres]"        # asyncpg + pgvector adapters
pip install "arc-agentkit[redis]"           # redis-backed store
pip install "arc-agentkit[observability]"   # OpenTelemetry exporter
pip install "arc-agentkit[mcp]"             # Model Context Protocol client

Requirements¶

Python 3.12 or newer.
One of: an API key for an LLM provider (Anthropic, OpenAI, DeepSeek, OpenRouter are shipped as batteries-included presets), or the local claude CLI on your PATH (the ClaudeCliCognition plug-in delegates to it and needs no server-side key).

Verify the install¶

python -c "import agentkit; print(agentkit.__version__)"

Prints the installed version. If it errors, the wheel isn't on your Python path — check the interpreter your shell resolved.

Where to go next¶

Tutorial — fifteen minutes, five steps, one runnable script per step. Builds a real research-briefing agent with a tool, a stream, a budget, and a human-approval gate.
Landing — the four-theme grid (cognition · control · state · behaviour) that names every primitive. Start here for the mental model.
Why agentkit — twelve concrete guarantees and a side-by-side comparison with LangChain / LangGraph / Instructor / Pydantic-AI / hosted alternatives.
Cheatsheet — every primitive, tight code, skimmable in 90 seconds. Reach for it when you know what you want and just need the invocation.