Open source · Apache-2.0
AURA is an open source agent harness for production operations work, written in Rust and licensed Apache 2.0. It takes an LLM plus the tools you already run and turns them into an agent that can work a live incident: reading cluster state, querying logs and metrics, checking your runbooks, and reporting what it found.
You declare the agent in TOML. Workers are separate roles, each with its own prompt and its own list of tools, so the worker that inspects a cluster is not the worker that changes it. Tools arrive over MCP, which means anything with an MCP server is reachable without writing an adapter.
What it does
Model output alone does not make an agent you can point at production. The harness decides which tools each worker can reach, what pauses before a write, what happens when a tool call fails, what state survives across turns, and what record you have when the run is over. AURA is that harness, and it is open source. You own the config and the deployment, and the telemetry goes where you point it. Nothing has to route through a vendor for the agent to work.Use cases
Scope what the agent can reach
Every worker gets an explicit tool list, so read access and write access live in different roles across the agent.
Stop before a destructive call
Approval gates name the mutating tools that need a human decision. Reads run straight through; writes wait for a yes.
Use the tools you already run
MCP servers over HTTP, SSE, and stdio: Kubernetes, Prometheus, Grafana, GitHub, Notion, AWS, Mezmo, and anything else.
See what the agent did
Every model call and tool call emits an OpenTelemetry span in OpenInference conventions, readable in any backend.
Answer from your runbooks
RAG over in-memory, Qdrant, or Bedrock knowledge bases, plus skills that load task instructions only when needed.
Run it where your data lives
OpenAI, Anthropic, Gemini, Bedrock, OpenRouter, or a local Ollama instance. Same config, different provider block.
Start here
Local quickstart
Install the binary, run
aura init, connect an MCP server, and start asking questions.Docker quickstart
Bring up AURA and a demo environment with Compose, no local installs required.
Kubernetes quickstart
Run AURA as a long-lived in-cluster service with the Helm chart and a read-only Kubernetes MCP server.
Examples
Each of these is a recorded run against a live environment. Where a video has a written walkthrough, the link goes straight to it.Install and connect
Install AURA and debug an incident
Install on a desktop, run
aura init against an Anthropic model, wire in a Grafana MCP server, and investigate.Watch the walkthrough →Install in Kubernetes with Helm
Helm install AURA and the Kubernetes MCP server into a cluster, read-only through both RBAC and a startup flag.Watch the walkthrough →
Connect MCP servers and scope tools
Add Mezmo, GitHub over HTTP streamable, and Grafana over stdio, then narrow which worker holds which tools.
Wire providers and pick a model
Point AURA at OpenAI, Anthropic, Gemini, Bedrock, OpenRouter, or a local Ollama with one provider block.
Investigate an incident
Find an unnoticed crash loop
A one-sentence health check turns up a pod in CrashLoopBackOff with 788 restarts that nobody asked about.
Turn a Slack report into a root cause
A coworker’s message with no service name or time range. The log worker builds a window and traces checkout to payment.
Debug a Lambda timeout
AURA connects through the AWS MCP proxy with existing CLI credentials, finds the log group, and verifies the timeout.
Debug in Dash0 with no custom adapter
Two config sections point AURA at Dash0’s hosted MCP server, read-only, with credentials in environment variables.
Debug Docker container permissions
The agent doesn’t know the stack runs on Compose. It opens with generic advice, then works down to the mount and owner.
Trace a request across services
The tracing worker follows a slow request from the gateway through three services and back, no manual span-hunting.
Give it your runbooks
Read Notion runbooks during an incident
A read-only Notion key in
.env and a runbook worker, so on-call guidance comes from your pages instead of the model.Use GitHub as a runbook database
A fine-grained token scoped to one repo, plus sources of truth and routing rules for a runbook engineer worker.
Keep it contained
Add a human breakpoint
Deny a
pods_delete call and confirm nothing changed, then approve the same call and watch it run.Run entirely inside AWS with Bedrock
Bedrock as the provider, a managed knowledge base doing retrieval in-account, and every call landing in CloudTrail.
Let it run on its own
Schedule a Kubernetes guardian
A cron job starts a one-shot run with inspect access and exactly one mutation tool, bounded by prompt and by RBAC.
Fix a failed Argo CD deployment
A sync-fail hook sends a symptom over A2A. AURA finds the version mismatch and edits the one config map it may touch.
Go from log line to merged fix
A GitHub worker with a reduced tool list and an approval gate on issue creation, carrying a root cause into a PR.
Sweep for drift on a schedule
A nightly run compares live cluster state against declared config and opens an issue for each drift it finds.
Join the project
Join the AURA Slack
Ask questions, share what you are running, and talk through an approach before you open a PR.
Get started on GitHub
Apache 2.0, written in Rust. Clone it, read the config examples, and open an issue when something does not fit.
Star the repo
Stars are how other SREs find the project. If AURA saved you an afternoon, a star is the cheapest way to say so.
Have a use case in mind?
Tell us the tool, the failure you keep hitting, and what the agent should do. We build these into walkthroughs.