Le Répertoire des AgentsFR · EN · ES · JA · HI

“I want a personal assistant that's always on, in my messaging apps”

Technical users comfortable with the command line, who want an assistant reachable day and night where they already chat — and who accept the responsibility that comes with such powerful software.

Recommended agent: OpenClaw

💰 Free open-source software; you pay for the models plugged into it — often $10 to $30/month in light use, or $0 with local models. Level: Advanced 🌍 Austria Free, open-source software (MIT license), with no subscription; only the calls to the plugged-in models (or hosting) cost anything

Price verified on 2026-10-29.

OpenClaw (openclaw.ai), the 2026 phenomenon open-source personal assistant (several hundred thousand stars on GitHub): it installs on your machine or a small server, lives in your messaging apps (WhatsApp, Telegram, Discord, Slack, iMessage…) and acts continuously — files, browser, messages, automations — with the model of your choice, including local models via Ollama.

How it works

Trigger: You message it in your messaging app, or a scheduled task comes due — understands the request, uses its tools (files, browser, applications) and reports back in the conversation, and you get an assistant reachable everywhere, moving your tasks forward even with the screen closed.

Step-by-step setup

Install OpenClaw on a dedicated machine or a small server — not on your main workstation to start.
Run the setup wizard (openclaw onboard) and plug in ONE messaging app to test.
Connect your model provider (API key, pay-per-use pricing) or a local model via Ollama.
Add “skills” and access one at a time, auditing each: see the guardrails, they count double here.

⚠️ Guardrails: It's the most powerful tool in this directory — and the riskiest: its security history includes critical flaws fixed in 2026, and analyses found problems in a significant share of public “skills”. Keep it permanently up to date, run it in isolation (container, dedicated machine), install no unaudited skill, and never give it your banking access.

Its particular angle

Like every general-purpose assistant, it can chat, write, summarize, translate and help with code. What sets it apart: it runs at home — on your devices, under your control — rather than in a publisher's cloud.

🔧 Technical details

Installation: Self-hosted
Prerequisites: macOS, Linux or Windows; Node.js 24.22.1+ or 26.12.1+ (Node 26 recommended, the installer can provision a supported runtime); an API key or CLI login for the chosen model provider, or a local model server.
Hardware: No general CPU/RAM/storage minimum documented for the Gateway; the Docker documentation requires at least 6 GB of RAM to build the source image locally (prebuilt images avoid this requirement) and enough disk for images and logs.
Self-hosting: yes — self-hosted Gateway on your own machine or a server, including in Docker (repo setup script and Docker Compose); the official README states there is no paid tier, no hosted service, and no project token.
License: open_source — MIT License (LICENSE file and license mention in the official repo github.com/openclaw/openclaw).
API & integrations: API: Yes · SDK: plugin SDK (openclaw/plugin-sdk) · CLI: Yes · Webhooks: Yes · MCP: Yes
Models: Hosted and local providers interchangeable as plugins (the official README cites Claude and Codex as harnesses, and an API key for the chosen provider is required); personal keys (BYOK) or reused CLI login. Local models via Ollama: built-in plugin, native Ollama API, local discovery (e.g. ollama/llama3.3), local-only, Ollama cloud or cloud + local modes.
Data: State, memory and credentials live on the user's hardware. Prompts go to the configured model provider and messaging platforms, plus any diagnostics export the user enables; by default, OpenClaw only contacts the publisher for a daily version check, with anonymous statistics in opt-in.
Documented limits: “Personal assistant” security model: a single trusted-operator boundary per Gateway, with no hostile multi-tenant isolation on a shared Gateway. Tools run on the host for the main session unless sandboxing is configured.
Operations: Gateway JSONL file logs under /tmp/openclaw/openclaw-YYYY-MM-DD.log (rotation at 100 MB, up to 5 archives), Logs tab in the Control UI and `openclaw logs --follow`, configurable secret redaction, diagnostics and OpenTelemetry documented in the configuration reference, and `openclaw security audit`.

Official sources: openclaw.ai · docs.openclaw.ai · github.com · docs.openclaw.ai · docs.openclaw.ai · docs.openclaw.ai — verified on November 1, 2026.

Find your agent with the quiz →

Frequently asked questions

How much does OpenClaw cost?

Free open source software; you pay for the models plugged into it — often $10 to $30/month in light use, or $0 with local models. Free open source software (MIT license), with no subscription; only calls to the plugged-in models (or hosting) cost anything. Price verified on October 29, 2026.

Is OpenClaw free?

Yes. Free open source software (MIT license), with no subscription; only calls to the plugged-in models (or hosting) cost anything.

Is OpenClaw hard to get started with?

Level: Advanced. Setup takes 4 steps; the first: Install OpenClaw on a dedicated machine or a small server — not on your main workstation to start.

What limits and guardrails should I know about with OpenClaw?

It is the most powerful — and the riskiest — tool in this directory: its security history includes critical flaws fixed in 2026, and analyses have found problems in a significant share of public “skills”. Keep it permanently up to date, run it isolated (container, dedicated machine), don't install any unaudited skill, and never give it your banking access.