Skip to main content
Prerequisites: macOS or Linux, and an LLM API key (OpenAI, Anthropic, Gemini, Bedrock, or OpenRouter), or a local Ollama instance.

​
1. Install AURA

This pipes a remote script to bash; to review the script or its checksum behavior first, see the Installation page.
This gets you the aura CLI, which is all this quickstart needs. The script auto-selects the best install method for your platform: a native package, Homebrew, or a direct binary download. If aura isn’t found after a direct download, add the install directory to your PATH:
The script takes no arguments and is configured entirely by environment variables. See Installation for all install methods and environment variables.

​
2. Generate a config

The wizard asks a few questions, then writes a ready-to-run config.toml in the current directory:
  • Provider. It first checks your environment for a conventional API key (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY) and suggests the matching provider as the default.
  • API key. If that variable is already set, it asks whether to use it. Otherwise it prompts for the key with masked input.
  • Model. It queries the provider’s live model list and offers a short, curated shortlist. Pick by number, accept the default, or type any model id.
  • Output. It writes config.toml, plus a .env only when you entered a key that wasn’t already in your environment.
Your key never lands in config.toml. The generated file references it as api_key = "{{ env.OPENAI_API_KEY }}" instead. If a .env was written, add it to your .gitignore.

​
3. Start chatting

AURA launches as a local chat assistant, loading ./config.toml and running the agents on your machine.
The generated config is an SRE orchestrator: a coordinator that routes each request to one of three specialist workers (incident-responder, metrics-analyst, and log-analyst), running them in parallel when the tasks are independent. You can add or remove workers, rewrite their prompts, and change any other option in config.toml later, using the Configuration Reference.

​
4. Connect your tools

Right now AURA’s workers can reason about problems and analyze data you paste in, but they can’t query anything yet. At the AURA chat prompt, type:
This launches a guided wizard that adds an MCP server to your config, giving the workers tools they can call. Pick one from the built-in catalog: Or choose custom to configure any other MCP server over http_streamable, sse, or stdio. The wizard then:
  1. Collects credentials with masked input, or reuses an environment variable you already have set.
  2. Connects to the server and lists the tools it discovered, all in memory, before anything touches disk. A credential that doesn’t work never reaches your config.
  3. Previews the exact [mcp.servers.<name>] block and asks whether to write it. Pressing Enter accepts. If the connection check failed, the wizard says so and pressing Enter declines instead.
  4. Writes the config, then asks which workers should see the new tools and records an mcp_filter for each. You might give the Kubernetes tools to incident-responder alone and leave log-analyst without them.
Secrets go to a .env beside your config; the TOML gets {{ env.VAR }} placeholders. AURA resolves them only to authenticate to the MCP server, so your credentials never enter a prompt or a tool definition and the model never sees them. Nothing is written until you confirm.

​
5. Restart and put it to work

MCP servers are loaded at startup, so restart AURA to activate the one you just added. Run /quit, then:
Try the starter prompt the wizard printed for your server. For PagerDuty:
Who is on call right now, and are there any open incidents?
Run /mcp at any time to list the servers the active agent has configured.

​
What’s next

​
Troubleshooting

​
aura: command not found

The install directory isn’t on your PATH. Run export PATH="$HOME/.local/bin:$PATH" and add it to your shell profile.

​
aura init can’t list models

aura init fetches the provider’s model list to confirm your key works and to build the shortlist it offers you. If that request fails, because you’re offline or the key was rejected, it warns and continues, allowing you to type the model id yourself. Pass --offline to skip the request entirely. AWS Bedrock has no model-list endpoint, so it’s always skipped.

​
Workers say they have no tools

No MCP servers are configured yet. Run /mcp add, then restart aura.

​
/mcp add isn’t available

The wizard edits the config.toml on your machine, so it’s only available when AURA runs agents from that file. Connecting to an AURA web server with --api-url disables the /mcp add command, since the config lives on the server instead.

​
Newly added MCP server isn’t showing up

There’s no hot reload. Restart aura after /mcp add finishes.