Self-hosted AI assistant that you actually own
The privacy-first self-hosted AI stack for developers, teams, and tinkerers who refuse to hand their data to the next AI vendor. OpenClaw runs the agent on your hardware. MemoryClaw keeps it encrypted, backed up, and portable. Both are free to start.
Open source runtime. Encrypted backups. Free MemoryClaw plan.
The case for self-hosted AI
Hosted AI is convenient, but the convenience comes with hidden costs. Your data trains the next model. Your prompts feed someone else's analytics. Your bills change without warning. Self-hosted AI flips that equation.
Data sovereignty
Conversations, embeddings, and memory live on hardware you control. Compliance, GDPR, internal policy — all easier when nothing leaves your perimeter unencrypted.
Cost predictability
A self-hosted AI agent has flat infrastructure costs. No per-message billing surprises, no rate limits when your usage spikes, no enterprise upsell when you outgrow the hobby plan.
Model freedom
Swap providers, use multiple models, or run fully offline. OpenClaw is model-agnostic so you are never locked to a single vendor's roadmap or pricing.
Persistent context
Self-hosted means you can store as much memory as you want, structured how you want. No 30-day retention limits, no opaque deletion policies.
The self-hosted AI stack
Two open building blocks. One ten-minute setup.
OpenClaw
The open-source agent runtime. Hosts your prompts, tools, skills, and per-conversation memory. Connects to any LLM provider — local or hosted — via pluggable connectors. Runs on macOS, Linux, and Windows (WSL).
- •Model-agnostic by design
- •Plugin and skill ecosystem
- •Local memory store with structured threads
- •Auditable, open-source codebase
MemoryClaw
The encrypted backup layer. Compresses your OpenClaw directory, encrypts it with AES-256-GCM using your passphrase, and ships it to cloud storage you can restore from on any machine. We hold ciphertext, not plaintext.
- •Zero-knowledge encryption
- •Versioned backups with rollback
- •Multi-agent support
- •One-command restore on any OS
Self-hosted vs. hosted AI
The honest comparison. We are not pretending hosted AI has no advantages — but the trade-offs are real.
| Self-hosted AI (OpenClaw + MemoryClaw) | Hosted AI assistant | |
|---|---|---|
| Data location | Your hardware. Encrypted backups in your cloud storage. | Vendor servers. Retention determined by their policy. |
| Model choice | Any LLM. Switch freely. Local models supported. | Whatever the vendor offers, when they offer it. |
| Cost model | Flat infrastructure plus optional API usage. Free MemoryClaw tier. | Per-message or per-token. Spikes hit your bill. |
| Memory retention | Unlimited. You control the policy. | Vendor limits. Often opaque or auto-expiring. |
| Vendor lock-in | None. Both layers are portable. | High. Migration is rarely supported. |
| Disaster recovery | Encrypted MemoryClaw backups. Restore in one command. | Vendor dependent. Often nothing for the consumer plan. |
Set up the stack in two commands
No config files. No yaml. No "first set the environment variable".
1. Install the OpenClaw runtime
2. Install MemoryClaw to back it up
Both installers walk you through login and run a smoke test. You'll have a working self-hosted AI agent with encrypted backup before your coffee finishes brewing.
Neural Memory for self-hosted agents
MemoryClaw GNM (Git-Neural Memory) gives your self-hosted agent structured, encrypted, searchable memory beyond the file-system backup.
Your agent can store engrams — typed memory units it can recall by concept, not just keyword. The cortex consolidates related engrams over time, and recall surfaces the right context at the right moment. Everything is encrypted client-side before it leaves your machine; the server never sees plaintext, not even your search queries.
The Free tier includes 10,000 engrams + 1,000 recalls per month. The Dev plan ($9/mo) raises this to 50,000 engrams + 20,000 recalls + unlimited resolves and on-demand consolidation.
Read the Neural Memory docsPrivacy guarantees we make in code
End-to-end encrypted
We never see plaintext
Audit the runtime
Start without risk
Self-hosted AI FAQ
Honest answers to the questions developers ask before adopting a privacy-first AI stack.
What does self-hosted AI actually mean?
Self-hosted AI means the runtime, the model (or the bridge to the model), and the data all live on infrastructure you control. No SaaS dashboard, no third-party logging, no vendor lock-in. Open source agent frameworks like OpenClaw make this practical for individual developers, not just enterprises.
Do I need a GPU to run a self-hosted AI assistant?
Not necessarily. OpenClaw is model-agnostic — you can route to a hosted API like OpenAI or Anthropic and still keep all the agent state, memory, and configuration local. If you want fully offline inference, smaller models like Llama 3 or Mistral run on consumer GPUs and even some CPUs.
How is this different from running ChatGPT in incognito?
ChatGPT and similar hosted assistants log every conversation, retain training data rights, and operate inside a single vendor ecosystem. A self-hosted AI assistant on OpenClaw lives on your machine, gives you direct access to the memory store, lets you swap models freely, and remains operational even if a provider changes pricing or shuts down.
Why do I need MemoryClaw if my AI is already self-hosted?
Self-hosted does not mean indestructible. Disks fail, OSes get reinstalled, laptops get replaced. MemoryClaw gives your self-hosted AI the same disaster recovery story as a cloud service — encrypted, automatic backups that restore your full agent state on any machine — without surrendering control of the data.
Is the privacy claim auditable?
Yes. OpenClaw is open source and runs on your machine, so you can audit the runtime end-to-end. MemoryClaw encrypts every backup with AES-256-GCM using a passphrase only you know, before it ever leaves your machine. Our servers see ciphertext only — by design and by code path.
Can I switch between models without losing my agent's memory?
Yes. OpenClaw separates memory and model. Memory lives on disk and is portable across providers. You can move from a local Llama model to OpenAI to Anthropic without losing your agent's history or skill configuration.
Is the free tier enough for a personal self-hosted assistant?
For most personal setups, yes. The free MemoryClaw plan offers 24 backups per day and 500MB of encrypted storage — enough for daily backups of a single OpenClaw instance over many months. Heavier users can upgrade to the Dev plan at $9/month for unlimited backups, 5GB of storage, unlimited agents, and 50,000 Neural Memory memories a month.
Stop renting your AI assistant
Run it on your hardware. Back it up with MemoryClaw. Restore on any machine, any time. Free tier included.
Get Started FreeDev plan from $9/mo when you scale beyond personal use.