"Self-hosted" sounds like full control, but what stays on your server and what can still leave it depends on one architectural decision you make before you deploy.
Short answer: A self hosted AI personal assistant runs its orchestration, memory, conversation history, and documents on a Linux server you own or rent. When the assistant reasons about a complex task, that prompt travels to an external model gateway such as Anthropic Claude or OpenAI by default, unless you provision dedicated hardware running local model weights. The minimum server for the external-gateway configuration is 2 vCPU, 4 GB RAM, and 20 GB SSD.
How a self-hosted AI assistant actually works
Nothing about the runtime is shared with a vendor. Orchestration, task tracking, and workflows execute on your machine, and each agent runs as its own isolated system process with its own tool grants. One agent cannot read another agent's mailbox, and it cannot call a tool it was not given.
The boundary is easier to hold in mind as a list.
| What | Where it lives |
|---|---|
| Conversation history | On your server |
| Documents and memory | On your server |
| Credentials and API keys | On your server, never inside the model's reasoning context |
| Reasoning inference | External model gateway by default, local with dedicated CPU or GPU hardware |
| Channel messages (email, Telegram, voice) | Transit those external networks |
Credentials follow the same rule. API keys and mailbox passwords stay in a local secret store, and an agent resolves a secret reference through a local tool wrapper when it needs one. The value is used, not quoted, so it never appears in the text the model reasons over.
Every state change and agent action is written to an append-only event log on the server. That log is what lets you reconstruct what an agent did, and whether it received explicit human confirmation before acting, weeks later.
The trade-off is the hybrid inference model itself. Runtime, storage, and credentials stay in single-tenant ownership, while complex reasoning may cross a network boundary to a provider whose commercial zero-retention API terms govern that request. The context payload for that request leaves your server. The database, the documents, and the memory do not. For the same boundary across several agents, roles, and shared memory, self-hosted AI agents is the broader reference.
What a working setup looks like
A single agent covers the personal-assistant pattern. Give it a role, inbox manager for example, and a scope that names the tools it may call: read mail, search documents, draft a reply. Nothing outside that scope is reachable.
Its email credentials sit in the secret store. The agent reads the inbox over IMAP and drafts replies, and those credentials are never placed in the model's reasoning context at any point in that loop.
Reference material such as contracts, procedures, and standing instructions lives in a Git-backed document repository on the server, with revision history. When the agent works a task, it retrieves the document instead of reconstructing it from memory, and the retrieval stays inside the workspace it was granted.
Complex drafting and analysis go to the external frontier model gateway you configured. The runtime, the conversation history, and the documents stay where they are.
Before anything leaves the node, the platform holds it. Sending an external email, publishing content, deleting a record, running a privileged command on the server, and spending money all stop at a human-in-the-loop gate and wait for your explicit confirmation. That gate is enforced by the platform rather than by the agent's judgement, so a badly written prompt cannot switch it off.
Steps to reach a working personal assistant
- Provision a Linux server (Ubuntu or Debian) with at least 2 vCPU, 4 GB RAM, and 20 GB SSD. That is the floor for a node that routes reasoning to an external gateway. Local model weights need compute sized to the weights you choose.
- Obtain an API key for at least one external frontier model gateway: Anthropic, Google Gemini, or OpenAI.
- Install the Smart Node runtime following the prerequisites and installation runbook at /docs/install.
- Create an agent with a defined role and scope. Add its SMTP and IMAP credentials to the secret store during configuration, not in a configuration file and not in a prompt.
- Enable the human-in-the-loop gate for outbound email, content publishing, and record deletion, and verify it is active before the agent connects to a live inbox.
- Open the browser panel and start a conversation. Confirm the agent reads the inbox, drafts a reply, and holds for your explicit confirmation before it sends.
Common mistakes
Most setup problems trace to a mismatch between what the operator assumed and what was actually configured.
Warning: Running the runtime on your server does not mean reasoning stays on your server. Unless you configure the agent to use local model weights and provision hardware that can run them, prompts for complex tasks route to the external model gateway you configured.
Warning: The 2 vCPU, 4 GB RAM, and 20 GB SSD minimum applies to the external-gateway configuration only. Attempting to run even a small open-weight model on that specification will fail or produce latency you will not accept. Size the server to the weights you intend to run.
Warning: Installing the software does not confer regulatory compliance. GDPR, HIPAA, SOC2, and similar frameworks depend on infrastructure configuration, access policies, data retention rules, and documented administrative procedures that sit outside the platform.
Frequently asked questions
Does my data stay private if the AI uses an external model like Claude or GPT?
Conversation history, documents, and memory stay on your server. When a task routes to an external model gateway, the context payload for that request transits the provider's network under that provider's commercial zero-retention API terms. Full in-house inference requires local model weights on dedicated hardware, configured explicitly for that agent.
Can I run this without any internet connection?
Only when the node is configured with exclusively local model weights and no outbound connections to external model providers. In the default configuration, which uses an external gateway, the node needs internet access to that provider.
What hardware do I actually need?
Budget 2 vCPU, 4 GB RAM, and 20 GB SSD on a Linux VPS running Ubuntu or Debian if reasoning goes to an external gateway. Local model inference needs additional CPU or GPU compute sized to the specific model weights.
Will the assistant send emails on its own or does it ask me first?
It asks first. The platform enforces a human confirmation gate on every outbound email, content publication, and record deletion, and the agent waits at that gate for your explicit confirmation. It cannot bypass the gate on its own.
Does self-hosting make me GDPR or HIPAA compliant?
No. Installing the software does not confer compliance. GDPR, HIPAA, SOC2, and similar frameworks depend on overall infrastructure configuration, access policies, data retention practices, and documented administrative procedures. The platform provides technical controls (process isolation, a secret store, an event log, human confirmation gates) that can support a compliance architecture, but that architecture has to be designed and documented separately.
Where to go next
If your server meets the minimum and you know which gateway each agent will use, what remains is configuration. Set up your node.