A 24/7 local-first personal AI assistant framework powered by Claude Code / OpenAI Codex, featuring multi-channel messaging, long-term memory, a skill system, and task scheduling.
Agentara is a local-first 24/7 personal AI assistant framework that wraps Claude Code and OpenAI Codex through a managed session layer, extending their capabilities from terminal/IDE scenarios to everyday assistant use cases.
The architecture is organized into four layers: the shared layer defines cross-layer types and interfaces (AgentRunner abstraction, messaging channel interfaces, task payload types, etc.); the kernel layer handles core orchestration (Kernel singleton, SessionManager, TaskDispatcher, messaging gateway); the community layer provides concrete implementations (Claude runner, Feishu channel); the server layer exposes a RESTful API built on Hono. The BootLoader ensures directory integrity and safe lazy-loading of the Kernel.
The AgentRunner interface exposes only a stream() method returning an async iterator. The Session layer abstracts away runner differences with run()/stream() modes. SessionManager manages JSONL-persisted sessions with a no-concurrency, no-cache design. TaskDispatcher uses Bunqueue for session-internal serial (FIFO) and cross-session concurrent (default 4 slots) task scheduling.
Feature-wise, it supports a multi-channel messaging gateway (Feishu implemented, with streaming responses), Cron scheduled tasks, long-term local memory, 24 built-in skills (covering deep research, data analysis, content generation, dev assistance, etc.), file/image sending/receiving, and a React 19-based web management dashboard. All data is stored under ~/.agentara using SQLite + Drizzle ORM, with configuration validated via Zod.
Prerequisites are Bun (latest) and Claude Code or OpenAI Codex (at least one configured). After installation, run make dev to start — the backend API listens on localhost:1984 by default, and the frontend dev server runs on localhost:8000.
Unconfirmed items: README declares MIT license but no LICENSE file found in the repo; OpenAI Codex runner implementation details unconfirmed (only anthropic found in community directory); Hono API spec marked as TODO; no channels beyond Feishu implemented yet.