mezmo/aura

What is AURA?

AURA is an open‑source SRE‑agent platform that lets you run AI‑powered assistants (agents) inside your own infrastructure. It ships with a set of pre‑configured “specialist” agents that can query observability data, cloud resources, version control, ticket systems, etc., and use large language models (LLMs) to investigate and help resolve incidents.

---

Why it matters

• Production‑ready safety – All model calls, tool invocations and state are wrapped in configurable guard‑rails, approval steps and OpenTelemetry tracing, so you can meet security and compliance requirements.
• Provider‑agnostic – Works with OpenAI, Anthropic, AWS Bedrock, Gemini, Ollama, OpenRouter, or any model that speaks the OpenAI‑compatible API. Switching models is just a config change.
• Extensible & observable – Agents are defined in version‑controlled TOML, can call external tools, fetch RAG data from Qdrant or Bedrock KBs, and export full end‑to‑end traces for post‑mortems.
• Runs anywhere – CLI, Docker image, Helm chart, or embedded as a Rust library, even in air‑gapped environments.

---

Core concepts

Concept	What it does
Agent	A small AI‑driven worker with its own prompt, model, and tool set (e.g., a “Kubernetes inspector” or “PagerDuty responder”).
Agent Skills	Re‑usable instruction bundles that can be loaded on demand to give an agent domain‑specific knowledge.
MCP (Model Context Protocol)	A discovery protocol that lets AURA find external tools or RAG stores over HTTP, SSE or STDIO.
Guardrails / Approval	Policies that require human or webhook confirmation before a tool that could change state is executed.
Streaming API	Real‑time, token‑level output that can be consumed by chat UIs or other services.
OpenTelemetry export	Every LLM turn, tool call and orchestration step is emitted as a trace span for observability.

---

Getting started (quick recap)

1. Install – Run the provided install script (curl … | bash).
2. Initialize – aura init walks you through selecting an LLM provider and creates a aura.toml config.
3. Run – aura starts the local CLI chat; aura webserver runs it as a service; Docker/Helm options are also available.

---

Who should use it?

• SRE / DevOps teams that want AI assistance for incident triage, root‑cause analysis, and automated remediation while keeping data inside their own security perimeter.
• Platform engineers looking to embed LLM‑driven workflows into existing CI/CD or observability pipelines.
• Companies with strict compliance (air‑gapped, regulated) that need full control over model providers, data flow, and approval policies.

---

Limitations / things to watch

• AURA is a framework, not a turnkey AI model – you still need access to an LLM provider and to configure the appropriate tools/integrations.
• Security depends on how you configure the guardrails and secret handling; mis‑configuration could expose credentials to model prompts.
• The richness of the investigation depends on the quality of the specialist agents and the data sources you connect (e.g., Prometheus, Kubernetes, PagerDuty, etc.).

---

Where to learn more

• Documentation – https://docs.mezmo.com/aura (quickstarts, config reference, streaming API guide).
• Examples – pre‑built TOML configs in the repo (examples/complete/...).
• Community – Slack channel linked in the README for questions and roadmap discussion.

---

TL;DR

AURA is an open‑source, Rust‑based platform that lets you run AI‑powered SRE agents inside your own environment. It provides a safe, configurable, and observable way to hook LLMs into your monitoring, cloud, and ticket‑system tools, making incident investigation faster while keeping data under your control.