Self-Host Your Personal AI Assistant

An autonomous agent that lives in your messaging apps, manages your calendar and handles tasks. You decide what it can access and which language model powers it. Open source, on your server, under your control.

100% Made in GermanyGDPR compliantHourly billingNo minimum term
Telegram agent@ihr-seed
Guten Morgen, wie sieht mein Tag aus?
Guten Morgen. 3 Termine heute, der erste um 10:00 mit dem Designteam. 2 E-Mails als wichtig markiert, eine davon von Ihrer Steuerberaterin. Soll ich die Antwortentwürfe vorbereiten?
Ja, und verschieb den 14-Uhr-Call auf morgen.
Kalender E-Mail Gedächtnis Lokal verarbeitet
02

What a personal AI assistant handles for you

Morning briefing via messenger

A compact summary every morning: upcoming appointments, relevant emails and the weather. Delivered automatically at your preferred time, right where you already read. The assistant knows your priorities and filters out the noise.

Email triage and draft replies

The agent prioritizes your inbox, summarizes long threads and drafts replies in your tone of voice. You just approve instead of typing. Client inquiries, invoices and newsletters are sorted automatically.

Calendar management across all accounts

Check availability, send invitations and detect conflicts. Just say what you need. The rest happens in the background. Works through the same messaging app you already use.

Research and summarization

Search the web, summarize articles and documents, compare sources. Results land structured in your conversation. Useful for market research, competitive analysis or preparing for client meetings.

Automate team communication

Standup summaries, task tracking and reminders directly in your team channel. With workflow automation, more complex processes can be built where the agent handles individual steps.

03

One agent. All channels. Your control.

Messages flow from all platforms to the agent on your server. The agent thinks using your chosen language model and replies where the message came from. With cloud models, processed content is sent to the respective provider. For complete data sovereignty, a local model can be used instead.

Messaging ChannelsLanguage Models
TelegramBot API
WhatsAppBusiness
SlackWorkspace
DiscordGateway
Your Server
AI AgentPlans, decides, acts
MemoryPreferences, projects, contacts
ToolsCalendar, mail, browser, shell
ClaudeAnthropic
GPTOpenAI
GeminiGoogle
Local ModelsOllama

Messaging Channels

TelegramBot API
WhatsAppBusiness
SlackWorkspace
DiscordGateway
Your Server
AI AgentPlans, decides, acts
MemoryPreferences, projects, contacts
ToolsCalendar, mail, browser, shell

Language Models

ClaudeAnthropic
GPTOpenAI
GeminiGoogle
Local ModelsOllama
You decide which data the agent processes

The agent on your server decides which tool it needs for a task. With cloud language models, processed text leaves the server towards the provider. To avoid that, run a local model. This requires more hardware but offers complete data sovereignty.

04

Three tools compared objectively

OpenClaw 375,000 Stars

Strengths

  • Largest community and the broadest ecosystem
  • Over 22 messaging platforms out of the box
  • Model-agnostic: over 200 language models

Weaknesses

  • Large codebase with a corresponding number of security advisories, requires careful configuration
  • Very high release frequency (almost daily), guides become outdated quickly
  • Founder moved to OpenAI, project leadership in transition
Recommended
Hermes Agent 170,000 Stars

Strengths

  • Automatic skill creation from successful tasks, building on the GEPA prompt optimizer (ICLR 2026)
  • Lightweight: 1 to 2 GB RAM, SQLite-based, no external database needed
  • Simplest setup of all three tools, a single command

Weaknesses

  • Youngest project (February 2026), still pre-v1.0. Less hardened than established alternatives as a young project
  • Fewer messaging channels than OpenClaw (6 vs. 22)
  • Less documentation and third-party resources than OpenClaw
NanoClaw 29,500 Stars

Strengths

  • Only 3,900 lines of code, auditable in minutes
  • Container isolation by default, no public CVEs to date
  • Minimal resource requirements, runs on the smallest server

Weaknesses

  • Locked to Claude (Anthropic), no model switching possible
  • Smallest community of the three projects
  • Fewer features than OpenClaw and Hermes Agent
For getting started, Hermes Agent is the recommended choice: the simplest setup, a lightweight system and automatic skill creation. For maximum platform diversity, choose OpenClaw. For security and auditability, NanoClaw offers the leanest option.
05

What your AI assistant can do

Persistent memory

The agent remembers preferences, ongoing projects and contacts. Every conversation builds on the last instead of starting from zero. The longer you use the assistant, the better it knows your context.

Multi-channel presence

One identity across all platforms. Continue on Telegram what you started on Slack. Context is preserved. The agent is reachable wherever you already communicate.

Proactive actions

Briefings, reminders and alerts without being asked. The agent reaches out when something is important, not just when prompted. This fundamentally distinguishes it from a chat interface that only responds to questions.

Tool integration

Read and write calendars, send emails, search the web, execute shell commands. The agent does not just generate text but takes action. You decide which tools are available.

Model freedom

Claude, GPT, Gemini or a local model via Ollama. Choose the brain and switch anytime without rebuilding your assistant. Cloud models currently deliver the highest quality. Local models offer complete data sovereignty but require significantly more hardware.

Personality and boundaries

Tone, behavior and safety boundaries are configurable. From a terse secretary to a talkative sparring partner. Rules like 'never send money' or 'never delete files without confirmation' protect against unintended actions.

06

Why a dedicated server instead of running locally?

Dedicated Server

Best choice for security, availability and professional use.

  • Security isolation: a compromised agent cannot access your workstation, SSH keys or browser sessions
  • Available 24/7: the agent works even when your laptop sleeps, reboots or travels
  • Container sandboxing with Docker and dedicated firewall limits the attack surface
  • Reachable from any device: smartphone, tablet, desktop, via the same messengers
  • No performance impact on your workstation, dedicated resources for the agent
  • Monthly server costs (from approx. 6 EUR) in addition to API costs
  • Basic Linux knowledge and regular maintenance required

Local Machine

Suitable for experimenting and users without security requirements.

  • No server costs, only API costs
  • Simpler start without SSH and server administration
  • Offline operation with local models possible
  • Compromised agent has access to your entire machine: files, credentials, browser, network
  • Not reachable when the machine is off, sleeping or on the go
  • Impacts performance of your workstation
  • No container sandboxing by default, larger attack surface
An AI agent on a dedicated server is isolated from your workstation. If the agent is compromised, it cannot access your SSH keys, browser sessions or local files. On a local machine, a compromised agent has access to everything you have. At the same time, the server runs around the clock and is reachable from any device via messenger.
07

How to set up your AI assistant

01Step 01

Create a Seed

Choose a model with at least 2 CPU and 4 GB RAM. This is sufficient for getting started with cloud language models. If you also want to run a local model, you need at least 8 GB RAM.

02Step 02

Install the agent

Hermes Agent can be set up via Docker Compose in a few minutes. Connect a Telegram bot, configure the LLM provider, set up the personality. Our step-by-step guide walks through every command.

03Step 03

Personalize and secure

Define tone and behavior rules. Specify which tools the agent may use. Set up safety boundaries. From now on, your assistant works around the clock.

08

Step-by-step setup

Tested on a current dataforest Seed. Every command explained, accessible without an account.

Automation · approx. 35 minBeginner

Set up an AI assistant with Hermes Agent

Step-by-step guide to set up Hermes Agent on a dataforest Seed. Docker Compose with Caddy, Telegram bot, LLM configuration and security hardening.

Hermes AgentAI AssistantDockerTelegramAutomation
Open guide

View all guides →

09

Configure a seed

Billed hourly, no minimum term, no setup fee. One server for your AI assistant, its memory and all associated tools.

Entry

Beginner

CPU allocation based on availability
At least Intel Xeon Gold
NVMe SSD storage
3-way replication via Ceph
DDR4
Balanced disk performance
3,65 €
/ Month
from
0,005848 €
/ Hour

Standard

All-rounder

AMD EPYC Turin
At least 2.6 GHz
Up to 4.5 GHz
NVMe SSD storage
3-way replication via Ceph
DDR5
Increased disk performance
9,01 €
/ Month
from
0,014439 €
/ Hour

Performance

CPU-optimized

AMD EPYC Turin (High Frequency)
At least 3.3 GHz
Up to 5 GHz
NVMe SSD storage
3-way replication via Ceph
DDR5
Maximum disk performance, IOPS-optimized
12,26 €
/ Month
from
0,019639 €
/ Hour

All prices incl. 19% VAT

10

Why dataforest cloud?

Data sovereignty

Your data stays in Germany. All seeds run in certified data centers in Frankfurt. No data transfers to third countries, full GDPR compliance.

Deployed in seconds

Seeds are provisioned automatically. From configuration to a running server takes only seconds. No waiting, no tickets.

Hourly billing

You only pay for what you use. No minimum terms, no setup fees. Seeds can be created and deleted at any time.

Full control

Root access, public API and full transparency. You decide what runs on your seed. No vendor lock-in, no hidden restrictions.

11

Bevor Sie loslegen.

What is the difference between an AI assistant and an AI chat service?
An AI chat service is an interface for a language model. You ask a question, get an answer, the session ends. A personal AI assistant is an autonomous agent that runs permanently on a server, connects to messaging apps, maintains persistent memory and acts proactively. It uses a language model as its thinking engine but does not replace it. The difference lies in autonomy, memory and tool access.
Which messaging platforms are supported?
This depends on the chosen agent framework. OpenClaw supports over 22 platforms including Telegram, WhatsApp, Slack and Discord. Hermes Agent covers the most common channels: Telegram, WhatsApp, Discord, Slack, Signal and email. Telegram is the easiest to set up with all projects and the recommended starting platform.
What does it realistically cost to operate?
Two cost blocks: the server and the language model. The agent itself is resource-efficient and runs on a Seed starting at around six euros per month. On top of that come the LLM provider costs based on usage. With moderate use (50 to 100 queries per day), API costs typically range from 5 to 30 euros per month. Running a local model eliminates API costs but requires a more powerful server.
Do I need a powerful GPU?
No, as long as you use a cloud language model like Claude or GPT. The model runs at the provider, your server only coordinates. A GPU only becomes relevant if you want to run a local model for complete data sovereignty. Small local models also run on CPU only but deliver noticeably lower quality than large cloud models.
Does my data really stay on my server?
That depends on the chosen language model. Your agent, its memory, configuration and messaging connections run entirely on your server. However, when the agent uses a cloud language model (Claude, GPT, Gemini), the text to be processed is sent to the respective provider. The advantage over commercial assistants: you control exactly which data the agent is allowed to process in the first place, and no permanent user profile is built. For complete data sovereignty, run a local model.
How secure is a self-hosted AI agent?
An AI agent with access to messages, calendar and email has a relevant attack surface. Risks like prompt injection (manipulated inputs that trick the agent into unintended actions) are real. Security depends heavily on the chosen framework and configuration. NanoClaw uses container isolation by default and has the smallest attack surface. Hermes Agent is actively maintained and uses a defense-in-depth security model. Key measures: restrictive tool permissions, behavior rules in the personality profile, timely updates and a firewall that only opens necessary ports.
What happens if the agent does something wrong?
Real incidents show that AI agents can perform unintended actions. This is why behavior rules are essential: rules like 'never send money without explicit confirmation' or 'never delete files' limit the scope of action. Most frameworks offer a confirmation mode where the agent asks before taking critical actions. Start with restricted permissions and expand gradually.
Can I switch the language model at any time?
With Hermes Agent and OpenClaw, yes. Both are model-agnostic: you swap the API key and model name in the configuration, the agent, its memory and all integrations remain unchanged. NanoClaw is tied to Anthropic Claude. Model-agnostic frameworks allow choosing the right model per task: an affordable one for simple answers, a powerful one for complex research.
Do I need technical knowledge?
Basic Linux skills are required: connecting to a server via SSH and running commands in the terminal. Docker experience is helpful but not mandatory. Hermes Agent can be installed with a single command, making it the easiest entry point among the three compared tools. Our step-by-step guide explains every command.
How is this different from workflow automation?
A workflow automation connects services by fixed rules: 'when new email, then Slack message'. An AI agent makes independent decisions based on natural language and its context. Both complement each other: the agent can trigger workflows, and workflows can feed data to the agent.
How do I back up my assistant's data?
The agent stores its memory and configuration in a local database. A daily backup of this data is recommended. The dataforest Cloud offers automatic daily offsite backups as an add-on option. Additionally, it is recommended to independently back up the database to external systems.

Our cloud newsletter

Stay up to date and get valuable tips by subscribing to our newsletter.

Any questions?

Our experts are happy to help. You'll be surprised how fast we are.

Visit FAQsContact an expert
Background image