kagura-agent

A memory-backed autonomous AI agent. It combines a pluggable agentic brain, Kagura Memory Cloud, a Docker security membrane, and a Slack/Discord cockpit. Use it from the CLI with run, repl, serve, doctor, and setup.

The default brain is Claude through the Claude Agent SDK for Python. The optional kagura-brain backend adds Claude, Codex, and Ollama-compatible providers behind the same interface.

Status: alpha, package version 0.7.1. The v0.1–v0.7 pure-Python skeleton is implemented and tested with strict typing and at least 95% coverage. Live Docker, cloud credential, and chat-transport edges remain deployment-specific.

Why kagura-agent

• Persistent memory: decisions, failures, and progress survive individual runs.
• Real hands: filesystem, shell, Git, Docker, cloud APIs, and custom MCP tools.
• Bounded autonomy: credentials, mounts, and egress are granted per task.
• Operator control: long-running work can be launched, resumed, approved, or stopped from a CLI or chat cockpit.

Quickstart

You need Python 3.11 or newer, a Kagura Memory Cloud login, and a configured brain. The base package intentionally ships without a brain dependency.

python -m venv .venv
# Activate the virtual environment, then install the default brain:
pip install 'kagura-agent[claude]'

# Authenticate both required services:
kagura auth login
claude

# Verify the environment and run a task:
kagura-agent doctor
kagura-agent run "summarize the repository layout"

ANTHROPIC_API_KEY can replace the Claude subscription login. For the kagura-brain Claude/Codex/Ollama backend, chat transports, Docker mode, Windows activation, and first-run troubleshooting, see Getting started.

Main commands

Architecture at a glance

The session layer orchestrates tasks and checkpoints without depending on a specific model SDK. The selected brain owns its agentic tool loop; Kagura Memory Cloud provides durable context; the host-side membrane controls filesystem, network, and credential access.

The full rationale, provider boundary, capability model, security membrane, and cockpit protocol live in the design document.

Documentation

Development

git clone https://github.com/kagura-ai/kagura-agent.git
cd kagura-agent
pip install -e '.[dev]'
pytest
mypy

See CONTRIBUTING.md for the full development workflow.

Related repositories

License

Apache License 2.0 — © 2026 Kagura AI. See NOTICE for attribution.