Open-source. No lock-in.
Open code. You can self-host it, fork it, audit it. If the provider changes tomorrow, you change server — your agent keeps working. No closed platforms holding your workflow hostage.
OpenClaw is an open-source architecture for AI agents. It orchestrates the LLM you choose together with real tools, persistent memory and open protocols so the agent can carry out tasks, not just answer messages. This page explains what makes it different, why it beats using ChatGPT or Claude on their own, and how to use it without building it yourself.
OpenClaw is an AI agent architecture. Open-source. Built so the agent doesn't stop at giving back text — so it does things. It reads your email. It calls an API. It writes to your CRM. It books a meeting. It chains 20 steps without you supervising each one.
The idea is old in the academic literature (Brooks 1991, writing about subsumption architectures) and new in practical engineering (Anthropic and OpenAI published their tool use specs only recently). OpenClaw takes the best of both worlds: a solid orchestrator, a standard protocol for tools (MCP), persistent memory and model agnosticism. And it packages all of it as free software, with no vendor lock-in.
If you got here searching for openclaw, openclaw docker o openclaw telegram, you have probably already sensed why it is worth it. This page tells you why it is doubly worth it, and why, if you don't fancy bringing up Docker at 3 AM on a Sunday, we can hand it to you already running.
If you're thinking "I already do this with ChatGPT or with the Claude API", stop. You are not doing it. What you're doing is talking to a brain in a jar. OpenClaw gives that brain a body: memory, tools, the ability to act in your world. Let me explain:
It reads a question, returns an answer, forgets everything. It can't execute anything in your world. Ask it to "reply to this email" and it gives you the text of the email — but it doesn't send it. Ask it to "book a meeting" and it tells you what to write — it doesn't open your calendar.
Same LLM — but now with an orchestrator. It takes an instruction, breaks it into steps, picks tools, executes, checks, retries if it fails, remembers what you asked for before and reports back with the result. And it can act for you when you're not there.
Five pieces. An orchestration layer, a memory layer, a tool bus (MCP), a model router and the runtime. If you have already wrestled with a hand-built agent in LangChain, this will look familiar — but integrated and in production, not as fragile DIY.
There is no magic. There are architecture decisions taken well, taken in public, with code you can read. Each one solves a concrete problem that home-made or proprietary agents have. Here they are:
Open code. You can self-host it, fork it, audit it. If the provider changes tomorrow, you change server — your agent keeps working. No closed platforms holding your workflow hostage.
GPT-4, Claude, Gemini, local Llama, Mistral. Use the model that works best for each task — or the cheapest one for the trivial jobs, and the premium one only when you need it. Changing provider is an ENV_VAR, not a rewrite of the app.
OpenClaw speaks MCP (Model Context Protocol) natively. Every tool is a clear contract: input, output, permissions. Adding a new one means writing a function and declaring it. No mile-long prompts instructing the model to "pretend" it is executing things.
Episodic memory (what happened) + vector memory (semantic) + long-term preferences. Your agent remembers who Ana is, which projects you have running, the tone you prefer. Without you typing it in again every time. And you can inspect and edit it.
An LLM answers when you speak to it. OpenClaw can act when something happens: webhook, cron, incoming email. Your agent works while you sleep. PR reviews as they land. Reports on Monday at 9. Alerts if something drifts.
It runs on your infrastructure (or on a private one we arrange for you). Sensitive data can go to a local model. Every action ends up in the log: what it called, with which arguments, what came back. Auditable, restrictable, supervisable. It is not a black box.
An honest, technical comparison against the four things people usually wrestle with when they start thinking about an agent: a direct LLM API, a framework like LangChain, a closed platform or OpenAI's Assistants API.
| Capability | LLM API (bare GPT/Claude) | LangChain / LlamaIndex DIY | Closed platform (Lindy, Relevance, Zapier AI) |
OpenClaw |
|---|---|---|---|---|
| stable tool use | Raw tool spec | Yes, you maintain it | Yes, limited to the catalogue | Yes, via standard MCP |
| persistent memory | No | You build it | Yes, opaque | Yes, inspectable |
| multi-model | One provider | Yes, with effort | Whatever they decide | Yes, native router |
| real self-hosting | No | Yes | No (closed SaaS) | Yes |
| auditable code | No | Partial | No | Yes, open-source |
| scheduled execution | No | You build it | Sometimes | Yes, native |
| vendor lock-in | High | Low (you're in control) | Total | Zero |
| cost at scale | Pay per token × task × client | Infra + maintenance | Monthly subscription x users | Infra + tokens for the model you choose |
Three examples of real flows. Each one chains several steps, uses memory and calls tools. All you do is write in plain language.
It reads your inbox, separates the important from the noise, writes a summary and suggests draft replies for the ones waiting on your input.
Cron runs the agent. It cross-checks Stripe, Google Analytics and HubSpot, spots anomalies vs. last week and sends you a readable report over Telegram.
You message it on Telegram. It pulls the context out of an email thread, finds the contact, looks for free slots, proposes and books. All in a single conversation.
Here they are in order of technical effort. Which one you pick depends on how many engineering hours you have, how much control you want and whether you're the type to build it yourself or to have it built for you.
You clone the official repo, bring up Docker, paste in your API keys, configure webhooks, wrestle with the Telegram token. Total control. The bill: your weekend (and the next one, and the next one).
For senior devs with timeThey give you the server with OpenClaw already installed. The config, the integrations and the maintenance are still yours. ~$29/month.
If you only want to skip the infrastructureAnother host, cheaper (~$16/month), less support. Telegram, config and updates are still your job.
The low-cost optionWe give you everything: infrastructure + OpenClaw configured + Telegram already connected + updates + 24/7 AI support + humans in your language. You just use it. €9.99/month. Free demo + 14-day trial.
For people who want to use it, not maintain itWe are not a reseller. We are not a wrapper. We live with OpenClaw: we use it every day internally, we contribute to the project, we push its limits. What you get is the same official OpenClaw — with everything that normally takes you a week to set up already done.
You use hablo. Behind it lives OpenClaw. And in OpenClaw's brain, whichever model you want.
15 minutes with Marta. She shows you OpenClaw running on your own real case, answers your questions and, if it convinces you, switches on your 14 free days. After that, €9.99/month. Cancel whenever you like.
Free demo · Cancel whenever you like · Human support in your language
An open-source architecture for AI agents: orchestrator + memory + tools (MCP) + model router. It turns a bare LLM into an agent that carries out real tasks.
An LLM replies with text. OpenClaw turns that text into chained actions: it executes functions, calls APIs, reads and writes in your systems, keeps memory across sessions and can act on a schedule without you being there.
Put bluntly: ChatGPT tells you what to write in the email. OpenClaw writes it, checks it, sends it and updates your CRM — and lets you know when Ana replies.
Yes, OpenClaw is free software. You can clone the repo, inspect it, fork it, deploy it wherever you like. hablo uses the official project, not a private version.
Whichever you choose. OpenClaw is model-agnostic: GPT-4, Claude, Gemini, local Llama, Mistral. You can switch with one line of configuration or — if you're on hablo — by sending a message to the agent itself: "switch to the claude-sonnet model for analytics".
MCP (Model Context Protocol) is a standard that lets models use external tools in a consistent way. OpenClaw speaks MCP natively, so you can integrate any tool that exposes the protocol without custom glue.
In practice: connecting Gmail, your CRM, Stripe or an internal API is a standard piece, not a fragile patch.
Yes. You need a server with Docker, a Postgres database (+ vector extension), an LLM provider with an API key, and if you want Telegram, a BotFather bot with its webhooks configured.
Set aside a couple of afternoons, ideally a long weekend. And when you're done, yes, you will have to maintain it (updates, alerts, scaling, backups). That is why hablo exists — we already do it, and we charge what the server alone would cost you.
OpenClaw runs on private infrastructure (yours if you self-host, an isolated private one at hablo). Prompts and responses go to the LLM you have chosen — if you want maximum privacy, you can route to a local model (Llama, Mistral on-prem).
Every agent has auditable logs of which tool it called with which arguments. You can restrict permissions tool by tool.
It depends on the route. Self-host: server (~€5-15/month) + LLM tokens (on real usage, ~€0.50-1.50/day in average cases). OpenClaw Cloud: $29/month. MyClaw: $16/month. hablo: €9.99/month (infrastructure) + tokens separately — but with €10 free to start, a 14-day trial and everything in your language.
If you self-host, you fix it. If you're on hablo, we fix it before you even notice — we have staging environments where we validate every update and only promote to production when it is stable.
With a person. 15 minutes. Usually with Marta. She shows you OpenClaw working on your own real case. No chatbot, no AI in disguise. If it convinces you, we switch on your 14 free days.
Yes. No lock-in, no penalty. One click from your panel.
Book a demo with Marta. We show it to you live on your own case and, if it convinces you, we switch on your 14 free days.
14 days without paying · +€10 in free tokens · Human support in your language
We’ll give you your own AI agent for 24 hours, with €5 of AI credit so you can put it to real work: answering your email, writing that report, whatever you want to test. No card.
We only verify your email and your WhatsApp. Prefer to look around first? Join the Community, where we answer questions every day.