Agent field guidePersonal agent

OpenClaw
situationalOpenClaw Foundation · beta · osi-open-source
Open-source personal agent that runs on your devices and works through chat channels, tools, skills, and schedules.
Visit openclaw.ai →Open in NoemiumSource verified 2026-08-19
ledger entry · observed by @whysanesanders
OpenClaw is the category reference for a self-controlled personal agent and has unusually serious security documentation. The tradeoff is operational: the main session can run tools on the host, so pairing, allowlists, sandboxing, secret isolation, and narrow service accounts are part of the product rather than optional cleanup.
layer
Personal agent
maturity
beta
deployment
local · self-hosted · hybrid
evidence
Source verified
Install
command · macOS, Linux, or WSL2
curl -fsSL https://openclaw.ai/install.sh | bashofficial install source →Best for
- A personal agent reachable from Telegram, WhatsApp, Slack, Discord, or other chats
- Scheduled work and cross-channel context on an existing computer or VPS
- Operators who want an extensible local control plane instead of a managed employee product
Requirements
- Node.js 22.22.3+, 24.15+, or 25.9+ when managing Node manually; Node 26 is recommended
- macOS, Linux, Windows, Docker, or another documented deployment path
- At least one supported model provider, local endpoint, or subscription login
Real cost scenarios
Existing computer and existing model access
$0/mo
No incremental hardware or model charge; electricity and existing subscriptions excluded.
Small VPS plus light API use
$5–$50/mo
Small always-on VPS and light-to-moderate paid inference; actual token use varies.
Local model
$0+/mo
Software is free; hardware purchase, electricity, and maintenance are not included.
Models and providers
OpenAI and ChatGPTAnthropicGooglexAIOpenRouterlocal and OpenAI-compatible endpoints
Channels and surfaces
WhatsAppTelegramSlackDiscordSignaliMessageGoogle Chatlocal dashboard and TUI
Security boundary
critical privilegeLocal Gateway and configured devices, plus every selected model provider and messaging channel.
- Pair unknown direct-message senders and keep channel allowlists narrow.
- Main-session tools run on the host unless sandboxing is explicitly configured.
- Isolate credentials and review third-party skills and plugins before installation.
Known limitations
- The large integration surface increases setup and maintenance work.
- Always-on behavior requires an always-on machine or server.
- Cost varies with model routing, channel APIs, and tool use; one sticker price is misleading.
- A local runtime does not keep data local when remote models or channels are enabled.
Typed evidence
- 1Availabilityopenclaw.aichecked 2026-08-19
- 2Installdocs.openclaw.ai/start/getting-startedchecked 2026-08-19
- 3Requirementsgithub.com/openclaw/openclawchecked 2026-08-19
- 4Pricingdocs.openclaw.ai/reference/token-usechecked 2026-08-19
- 5Licensegithub.com/openclaw/openclaw/blob/main/LICENSEchecked 2026-08-19
- 6Security / threat modeldocs.openclaw.ai/security/THREAT-MODEL-ATLASchecked 2026-08-19