AgentAura Web Quick Start

Connect a Linux server you already run, choose how the agent authenticates to its model, and deploy — with the exact fields and commands the flow asks for.

Web Quick Start

The web flow installs an AI agent runtime onto a server you already have. It does not create, sell or manage that server.

That is the whole shape of it, and it is worth being blunt about: the provider picker is gone from the web deploy flow. There is no "choose Hetzner, pick a size" step. You bring an address and credentials; AgentAura installs onto it over SSH.

Native app builds may offer different deployment options than the web. This page describes the web flow only. Follow whatever the app in front of you actually shows.

What you need first

A Linux server you administer. Debian or Ubuntu family — the installation uses apt and installs .deb packages, so a distribution without them will not work. Ubuntu 22.04 and 24.04 are the versions this is built and run against.

SSH access as an administrator. You will provide the address, the user, and either a password or a private key. A non-default port is supported.

Outbound internet from that server. Inbound ports are generally not required: the deployment establishes a Cloudflare Tunnel, so the gateway is reachable without opening anything in your firewall. If you find yourself opening inbound ports to fix a connection problem, check first whether your deployment uses a tunnel — opening them adds exposure without addressing the cause.

Credentials for a model. Either an API key, or a subscription login. Details below, because this is the step people get wrong.

You remain responsible for the server: its bill, its SSH access, its firewall, its operating-system updates, storage, backups and lifecycle. AgentAura installs software onto it and does not take any of that over.

Sizing

There is no fixed minimum published here, because the honest answer depends on which agent you run and what you ask it to do. A small instance is enough to try it; an agent that builds software wants more memory than one that answers questions. Start small — you are the one who can resize it.

Open AgentAura and sign in. Complete any verification shown before starting a deployment.

Step 2 — Decide how the agent authenticates

This is the choice that matters most, and there are two different kinds.

An API key

You pay the model provider directly for usage. Currently offered in the web flow:

Provider	Key format
Anthropic (Claude)	`sk-ant-…`
OpenAI (GPT)	`sk-…`
DeepSeek	`sk-…`

Instead of an API key, the agent signs in with an account you already pay for.

Claude Code — uses an OAuth token from your Claude subscription rather than an API key. Generate it on your own machine:
```
claude setup-token
```
It prints a token beginning sk-ant-oat01-…. Paste that into AgentAura. This is the default path for Claude Code deployments.
Codex — uses your ChatGPT account. Paste the full contents of ~/.codex/auth.json from a machine where you have already signed in to Codex.

Both of these are account credentials, not scoped API keys. Treat them the way you would treat the password to that account, and regenerate rather than share them.

The API-key field deliberately starts empty every time and is never pre-filled from a saved value. A stale key silently redeployed is worse than typing a fresh one. Keys can also be set after deployment.

Step 3 — Enter the server

The deploy form asks for:

Field	Default	Notes
IP address	—	The server's public address
SSH user	`root`	Must be able to install packages
Authentication	Password	Or paste a private key instead
SSH port	`22`	Change if yours differs

Servers you have used before can be saved and re-selected rather than retyped.

Step 4 — Deploy, and watch it

Start the deployment and read the output as it runs. It reports each step and any failure. This is the single most useful habit here: a deployment that fails at step three tells you why, and a deployment you did not watch just looks broken afterwards.

When it finishes, open the agent from a client and confirm it responds. A completed install that no client can reach is not a completed deployment.

Adding more agents

One account maps to one subdomain. Deploying another agent to a server you have already set up adds it to that server and reuses the existing subdomain and tunnel, rather than provisioning anything new.

Plans

Plan	Price	Entitlement
Free	$0	1 private agent
Pro Monthly	$9.99/month on the web	Up to 50 agents
Pro Annual	$95.88/year on the web	Up to 50 agents

App-store pricing differs from the web. The checkout or purchase sheet in front of you controls the final price and billing period. Server hosting and model usage are separate costs, billed to you by those providers — they are not part of any plan here. See pricing.

If it does not work

The deployment fails partway. Read the step it stopped on. The most common cause is SSH: wrong port, a user without install rights, or key authentication where the server expects a password.

It installed but no client can reach it. Check the tunnel rather than your firewall. Opening inbound ports is almost never the fix here.

The agent responds with an authentication error. The model credential expired or was rejected. Subscription tokens in particular do expire — generate a new one with claude setup-token and set it again.

See the FAQ and the server reference for more.

Installation Server & Deployment