Smart Node runs AI workflow automation on a Linux server you own or rent, so the orchestration layer, task history, and business documents stay on your infrastructure, not on a shared vendor platform.
Short answer: Smart Node deploys a multi-agent runtime on your server. Each agent handles a defined business function, executes tasks through explicit tool grants, and holds for your confirmation before sending external messages, deleting records, or spending money. Licensing is per node, not per seat: one server covers a team of any size.
Get started with a node you can inspect before you move your first process onto it.
Is this the right fit?
AI automation that touches real systems raises three questions at once: where the data sits, who approves what the agent does, and what the operation costs as the team grows. The lists below answer those before you read any further.
It fits when:
- Your organization handles internal data that cannot sit on a shared SaaS platform, and you need it to stay on hardware you control.
- You want agents to take real actions through your systems, but a person must approve anything that leaves the company or removes a record.
- Per-user licensing has become a scaling problem and a flat cost per node suits your budget better.
- You already run a Linux server on Ubuntu or Debian, or you can provision one with 2 CPU cores, 4 GB RAM, and 20 GB of SSD storage.
It does not fit when:
- Nobody on your team can administer a Linux server and no managed service provider handles that for you.
- You need a working deployment this week with no infrastructure work at all.
- You need full air-gapped operation with zero external network calls and you have no dedicated GPU hardware to run local model weights.
What you get
| Capability | What it means for you |
|---|---|
| Agents with assigned roles | Each agent runs as an isolated process with a declared list of tools. It can only use what you granted it, and it cannot reach credentials it was not authorized to use. |
| Workflows and an append-only event log | Tasks run in deterministic sequences, and every action and state change is written to a log on your server, so you can reconstruct what an agent did and when. |
| Human approval gates | Before an agent sends an external email, publishes content, deletes a record, runs a privileged system command, or authorizes spend, it stops and waits for a person to confirm. Nothing proceeds without that confirmation. |
| Communication channels | Web chat in the office panel, internal mail between agents, SMTP and IMAP email, Telegram bots, and voice interfaces. |
| Documents and memory on the same server | Documents live in a Git-backed repository with full version history. Knowledge retrieval uses vector-indexed memory with workspace isolation, so an agent retrieves only what its scope covers. |
| Model routing per agent | Each agent is configured for either a local open-weight model on your hardware or an external frontier gateway (Anthropic, Google Gemini, or OpenAI). Different agents in one deployment can run different models matched to the weight of their work. |
Warning: When a task routes to an external frontier model gateway, the context payload for that request leaves your network and travels to that provider under its commercial API policies. The runtime, task history, and stored documents stay on your server. The security architecture covers this boundary in full.
Set up your node and configure model routing the way your data policy requires.
What it costs to run, and the trade-off
Minimum server: 2 CPU cores, 4 GB RAM, and 20 GB of SSD storage running Ubuntu or Debian. A standard entry-level VPS meets this specification when reasoning routes to an external model gateway. Running all inference locally requires dedicated CPU or GPU compute sized to the model weights you choose, and the minimum VPS is not enough for high-accuracy local inference.
Licensing: one flat cost per node, with no per-user fee. A single node serves your whole team regardless of headcount, which is where per-seat licensing elimination shows up in the budget.
Ongoing costs: hosting or server rental, operating system updates, monitoring, and backups are yours or your provider's to run. Per-token charges accrue on the external model gateway account when complex tasks route there, at rates set by that provider.
The trade-off is straightforward: you swap a predictable SaaS subscription line item for infrastructure hosting and administrative responsibility. The data residency benefit is real for stored documents, task history, and configuration, and the boundary is the inference request itself, where a complex reasoning task sends a context payload to whichever external gateway is configured. If your organization cannot accept any prompt data crossing an external network, you must provision local GPU hardware and accept the accuracy and throughput of the model weights that can run on it.
Frequently asked questions before you decide
Does my business data go to OpenAI or Anthropic when my agents run tasks?
Only the context payload of a request you route to an external gateway leaves your server, and stored data does not. Your documents, task history, message threads, and memory stay in local storage on your hardware. Prompts are sent only when an agent is configured to use an external model and a task actually reaches it. If you need nothing to leave the host, configure local model inference and accept the hardware requirement that comes with it.
Can an agent send an email or post something publicly without asking me first?
No. External messages, published content, record deletions, privileged system commands, and monetary spend all sit behind a human-in-the-loop gate and wait for explicit confirmation. The agent prepares the action, you approve or reject it, and only then does it execute.
We have a 20-person team. Do we pay per user?
No. Licensing is per node, so twenty users cost the same as two. You pay for the server you run and for the model usage your tasks generate, not for seats.
Can I run this on a standard cloud VPS, or do I need a dedicated server?
A standard VPS is enough when reasoning routes to external model gateways: 2 CPU cores, 4 GB RAM, and 20 GB of SSD storage on Ubuntu or Debian. Dedicated hardware becomes necessary only when you want fully local inference without external API calls, because local model weights need CPU or GPU capacity matched to their size.
What technical skills does my organization need to deploy and operate this?
Linux administration. Deploying and running the node, keeping the operating system updated, monitoring the host, and managing backups are the operator's responsibility, whether that is your team or a managed service provider. The platform handles agent isolation, tool permissions, and the approval gates, but it does not remove the work of running a server.
Can I give different agents different AI models depending on what they do?
Yes. Model configuration is set per agent, so a routine agent can run on a small local model while a research or drafting agent routes to a frontier gateway. The hybrid inference model lets you match cost and capability to each role instead of paying frontier rates for every task.
How to start
- Provision a Linux server on Ubuntu or Debian with at least 2 CPU cores, 4 GB RAM, and 20 GB of SSD storage. You finish this step with a target host.
- Follow the installation runbook at read the install runbook to install the runtime and configure your first node. You finish this step with a node that is running and reachable.
- Define your first agent role, assign it a workflow, and connect a channel. You finish this step with an agent handling tasks and holding for your confirmation before any external action.
Before you commit, the component overview at what you get lists every part of the platform, and the security architecture documents the data boundary and credential handling in detail.
Deploy a private node when you have a host ready.