AI Agents Platform

An AI agents platform that runs on your server, not someone else's

Smart Node is a self-hosted AI agents platform for business operations. Multi-agent orchestration, no per-seat fees, and your data stays on your server.

Smart Node is a self-hosted AI agents platform for business operations: you deploy it once on a Linux server you own or rent, and it runs multi-agent workflows for your entire team without per-seat licensing.

Short answer: Smart Node runs a complete multi-agent orchestration runtime on a Linux server you control, with a minimum configuration of 2 vCPU, 8 GB RAM, and 20 GB SSD when agents use external model gateways. Your runtime, task queue, data, and credentials stay on your own infrastructure, and one deployment serves an entire team with no per-user licensing cost. The trade-off is operational responsibility: you or your administrator maintain the host.

The two costs that grow with a SaaS agent platform

Per-seat pricing looks reasonable at five users. At forty it is a line item that tracks the org chart rather than the work the agents do. The bill scales with headcount, not with tokens consumed, tasks completed, or CPU time used, so it rises every time you hire, whether or not that person ever touches an agent.

The second cost is quieter. Conversation history, task logs, and the business documents your agents produce sit on the vendor's servers under the vendor's retention and access policies. Mailbox passwords and third-party API keys that agent tools need may pass through or be held by infrastructure you do not administer. For a team handling customer correspondence, supplier terms, or internal records, that is a governance question rather than a preference.

The two costs compound in one more way: workflows built inside a SaaS agent platform do not come with you. Moving to another platform means rebuilding them.

What the platform includes

Smart Node brings AI agent orchestration, memory, documents, channels, and oversight into one deployment. Each row below names a component and what it changes for you.

ComponentWhat it means for you
Multi-agent orchestrationAgents coordinate through a built-in Message Bus, a task tracker, and a deterministic Workflow engine, with every state change written to an append-only Event log. No external orchestration service sits in the path.
Human oversight and process isolationBefore an agent sends an external email, publishes content, deletes a record, runs a privileged shell command, or commits spend, it stops for explicit human confirmation. Each agent otherwise runs in its own isolated Linux system account with declared capability limits.
Memory and DocumentsKnowledge accumulates in a local vector store with workspace isolation and audience tagging. Records and structured content live in a Git-backed document repository with version history and discussion threads. Neither is used to train third-party public models.
Communication channelsAgents connect to browser web chat, internal agent mail, SMTP and IMAP email, Telegram bots, and voice interfaces from the first deployment, and each conversation maps to a persistent thread.
Hybrid model routingEach agent can be configured with a different model. A lightweight task agent can run on the same VPS while a reasoning-heavy agent routes to an external frontier gateway, both on one node.
No per-seat licensingOne deployed node serves the whole team. The licensing cost does not rise with headcount.

If that inventory covers what your operation needs, Set up your node and choose the server it will run on.

The trade-off

What you take on is the managed layer. Host security, operating system updates, monitoring, backups, and recovery belong to your organization or to the integrator it engages, and nobody does that for you. The hybrid inference model has a boundary worth reading twice: runtime, storage, and credentials stay on your server, but when an agent routes a complex task to an external frontier gateway such as Anthropic, OpenAI, or Google Gemini, the context payload for that request crosses the network to that provider under its commercial zero-retention API policy, and messages sent through Telegram, public email networks, or cloud voice traverse those services as well. Egress to external gateways is explicit and metered, so you can see what leaves. Running with no external API calls at all is possible, but it requires dedicated CPU or GPU capacity sized to the model weights, beyond the base 2 vCPU and 4 GB. Per-seat subscription fees are replaced by infrastructure hosting and per-token gateway fees, and deploying and administering the node requires Linux proficiency or an integrator who has it.

Who this fits

It fits when:

  • Someone on your team, or an integrator you engage, can administer a Linux server.
  • Per-seat AI agent subscriptions are growing with headcount and have become a real line item.
  • Conversation history, task records, documents, and credentials have to sit on infrastructure you own or rent.
  • Workflows need a human confirmation gate before anything irreversible happens.
  • Different agent roles call for different models, a small local one for classification and a frontier gateway for hard reasoning.
  • Headcount makes flat infrastructure economics better than per-seat licensing.

It does not fit when:

  • Nobody in your team or support network can run a Linux server, and no integrator is engaged.
  • You need fully air-gapped operation with zero external network egress. The default hybrid configuration routes complex reasoning to external gateways.
  • You expect regulatory compliance such as HIPAA, GDPR, or SOC 2 out of the box from installation alone. Compliance depends on your infrastructure configuration, access controls, retention practices, and administrative procedures.
  • You are a single user for whom a SaaS tool is technically sufficient and cost is not the deciding factor.

If the first list describes your situation, Deploy a private node.

Frequently asked questions

Does my business data stay on my server if my agents use Claude or GPT?

The data at rest does. Runtime, storage, conversation history, task records, and credentials stay on your server. When an agent routes a task to an external frontier gateway, the prompt payload for that one request crosses the network to the provider under its commercial zero-retention API policy. Data held in local databases, the document repository, and the memory store is never used to train third-party public models.

Is there a per-seat fee, and what does it cost to run?

There is no per-seat licensing fee, and a single node serves an entire team regardless of headcount. Running costs are the server or VPS you host it on, plus per-token fees when agents call external model gateways. Fully local inference removes the gateway fees but requires hardware sized to the model weights. Software acquisition terms are separate from these running costs and are worth confirming for your deployment.

What stops an agent from sending emails or deleting records without asking me?

Human-in-the-loop gates do, and they sit at the action level rather than in an agent's instructions. Before an agent sends an external email, publishes content, deletes a record, runs a privileged shell command, or commits monetary spend, the platform holds the action and requires explicit human confirmation. Those categories are structural constraints, not a setting that can be switched off.

Can I run models entirely locally with no external API calls?

Yes, provided the server has the CPU or GPU capacity for local open-weight inference. The base specification of 2 vCPU, 8 GB RAM, and 20 GB SSD covers a node that uses external model gateways. Local inference needs dedicated compute sized to the chosen model weights, and both modes can run on the same node at the same time.

What server do I need to get started?

Ubuntu or Debian on a Linux VPS or dedicated host with at least 2 vCPU, 8 GB RAM, and 20 GB SSD when agents use external model gateways. Local model inference raises those numbers. The full prerequisites and installation runbook are at install prerequisites.

How is this different from building on an open-source agent framework?

An open-source agent framework gives you a library, while Smart Node gives you a running platform. Multi-agent coordination over a built-in Message Bus, task tracking, deterministic workflows, persistent memory, a Git-backed document repository, channel adapters for email, Telegram, web chat, and voice, human oversight gates, credential management, and process isolation are operational from deployment. Assembling the equivalent on top of a framework is a separate engineering project.

How to start

  1. Review the component overview to confirm the architecture matches your use case. You finish with a scope: which agents, which channels, which models.
  2. Check the install prerequisites and confirm your server meets them. You finish with a host that is ready for the deployment.
  3. Deploy the node on that server. You finish with a running multi-agent node on hardware you control.

When the fit is clear, Get started to deploy your node.

See Smart Node set up for your business.

Tell us about the leads you miss and the work that piles up. We run a live demo on your actual workflow, then you decide if the crew earns a place.