Aura: New Go-Based Self-Hosted AI Agent Launches

Overview

A new self-hosted AI agent platform named Aura has emerged in the open-source community, targeting users who prioritize privacy and infrastructure control. Written in Go, the project distinguishes itself as a local-first, provider-neutral environment designed for ongoing autonomous work. Rather than relying on a single cloud API, Aura allows operators to host the runtime, tools, and memory systems entirely on their own hardware.

The platform addresses a growing demand for AI agents that can operate continuously without leaking sensitive context to third-party servers. By decoupling the agent framework from any specific large language model (LLM) provider, Aura supports a wide array of inference options, ranging from cloud-based APIs to fully local deployment using tools like llama.cpp or Ollama. This flexibility ensures that users can maintain strict data sovereignty while still accessing cutting-edge reasoning capabilities.

Architecture and Deployment

Aura is engineered as a comprehensive suite rather than a simple chat interface. The core Go binary acts as the central nervous system, hosting the agent loop, a command-line interface (CLI), a Telegram gateway, and an embedded web cockpit. This monolithic yet modular design simplifies distribution while allowing for complex multi-user interactions.

For deployment, the project leverages Docker Compose to orchestrate a robust stack of supporting services. The default configuration spins up PostgreSQL for authoritative conversation storage, ArcadeDB for graph-native memory and full-text vector indexing, and Garage to function as an S3-compatible object store. Additionally, the stack includes bundled integrations for web search via SearXNG, document ingestion, voice processing, and embedding generation.

Hardware requirements are notably modest, making the platform accessible to hobbyists and small enterprises alike. According to recent benchmarks, a mini PC equipped with 16 GB of RAM is sufficient to run the default stack, which measured approximately 7 GB of memory usage with speech-to-text and text-to-speech services active. The system is designed to run efficiently without a local LLM by default, routing inference to the user’s chosen provider to optimize resource allocation.

Key Features and Capabilities

The agent loop within Aura is built around a streaming architecture that enforces shared step and time budgets, preventing runaway processes and ensuring controlled execution. A standout feature is its graph-native memory system, which stores facts, sources, and validity windows directly in ArcadeDB. This allows the agent to traverse temporal paths and return supporting evidence for its assertions, creating a more verifiable and contextual form of recall compared to traditional vector databases.

Tooling and Security

Aura provides extensive tooling capabilities, including a full host terminal and filesystem access. To mitigate risks associated with such powerful operations, the platform implements destructive-command approval gates and automatic secret redaction. The system also supports deferred tools and dynamic schema loading, enabling it to discover and utilize tools from mounted MCP (Model Context Protocol) servers on demand.

Security is further enhanced through a per-identity sandbox mechanism. When enabled, each user operates within a gVisor-based isolation layer on native Linux systems. This ensures that agents cannot escape their designated environment, and file deliverables are transferred over secure channels rather than exposed via direct filesystem paths.

Multi-User and Multi-Channel Support

The platform supports multi-user environments with distinct identities, each maintaining their own isolated memory database, workspace, and sandbox. Authentication is handled through Authula, which supports password-based login and optional TOTP verification for enrolled accounts. An admin audit view provides oversight into capability grants and system activity.

Users can interact with Aura through multiple channels, including a CLI REPL, the Telegram interface (supporting voice, photos, documents, and human-in-the-loop approvals), and a responsive web cockpit. The web interface utilizes AG-UI/SSE protocols, enabling mid-turn steering of agent activities and real-time adjustments to settings.

Model Integration and Extensibility

While DeepSeek-V4 Flash via OpenRouter is set as the default inference route, the cockpit settings allow for hot-reloading of model profiles without requiring a system restart. Users can switch between OpenRouter, ChatGPT plans, or local inference engines at will. This adaptability ensures that the agent can leverage the most cost-effective or privacy-preserving model for any given task.

The platform also emphasizes self-extension through skills and integrations. Operators can author custom skills or utilize bundled modules for productivity and information management (PIM), email, calendar, and even unofficial WhatsApp connectivity. The inclusion of a Studio within the web cockpit expands the agent’s utility to include image and video generation, editing, and multi-track video production.

Development practices within Aura reflect a high standard of quality assurance, with test coverage reportedly exceeding 85% across owned surfaces. The continuous integration pipeline executes unit, property, race, leak, mutation, and live integration tests, alongside browser-based validation via Playwright. This rigorous testing regimen aims to ensure stability for users deploying Aura in production-like environments.