Vote in Free Arena

Always-on AI agents compared New

Always-on agents keep working when your PC is off, connect to your apps and ask before risky actions such as paying or sending email. Here are the big ones and 8 free ones, with a reliability index and only what their makers state.

Agents compared135 from big labs, 8 free and open source
Usable in Italy today8all self-hosted; the big labs are US-first or unclear
Most reliableOpenClaw90/100 on our index
Checked on6 November 2026official pages and GitHub

Which agent fits you

Move the sliders to say what matters to you. Each agent gets an affinity score from the facts in its card below: Yes counts 1, Partly 0.5, No or Not stated 0, and Reliability is its index out of 100.

Compare the agents

Show:

13 agents shown

OpenClaw

OpenClaw (open source, MIT)Free

Your data and credentials stay on your hardware; runs as a background service.

Reliability index90/100
Generally available: Regular dated releases, latest v2026.11.1. Public for 11 months. 394,014 GitHub stars · 100+ commits in the last 30 days · 49 releases in the last 90 days. Stage 40/40 · Track record 20/30 · Upkeep 30/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyYour own computer or server; always on only while that machine is on
Watch the browser liveNot statedNot stated in the README
Custom MCP serversNot statedTools, skills and plugins; MCP not named in the README
Asks before risky actionsPartlyUnknown senders must be paired and approved before they can talk to it
Price
Free software; you pay only the model, or use a free or local one
Model
Any: hosted models (Claude, Codex) or local models
Where you talk to it
WhatsApp, Telegram, Slack, Discord, Teams, Signal, iMessage, Google Chat and 20+ more; apps for macOS, iOS, Android

Sources: openclaw/openclaw on GitHub

Hermes Agent

Nous Research (open source, MIT)Free

Built-in scheduler for daily reports and unattended jobs, memory, writes and improves its own skills.

Reliability index85/100
Early release or rolling out: Version 0.x, latest v0.23.0. Public for 15 months. 260,977 GitHub stars · 100+ commits in the last 30 days · 23 releases in the last 90 days. Stage 25/40 · Track record 30/30 · Upkeep 30/30
In Italy / EU nowYesRuns on your own computer, a server or a cloud sandbox, anywhere
Own computer, works with your PC offPartlyLocal, Docker, SSH, Modal, Daytona or Vercel Sandbox: always on if you run it on a server
Watch the browser liveNot statedNot stated in the README
Custom MCP serversNot statedNot named in the README
Asks before risky actionsNot statedNot stated in the README
Price
Free software; you pay only the model, or use a free one through OpenRouter or your own endpoint
Model
Any provider: Nous Portal, OpenRouter, OpenAI or your own endpoint
Where you talk to it
Telegram, Discord, Slack, WhatsApp, Signal, command line

Sources: NousResearch/hermes-agent on GitHub

ZeroClaw

ZeroClaw Labs (open source, Apache-2.0)Free

Small and fast, scheduled jobs and event-triggered procedures with approval gates.

Reliability index75/100
Early release or rolling out: Version 0.x, latest v0.8.7. Public for 8 months. 33,143 GitHub stars · 100+ commits in the last 30 days · 3 releases in the last 90 days. Stage 25/40 · Track record 20/30 · Upkeep 30/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyA single Rust program that runs as a background service on your machine or server
Watch the browser liveNot statedNot described in the README
Custom MCP serversNot statedNot named in the README
Asks before risky actionsYesDefault supervised mode: medium-risk actions need approval, high-risk ones are blocked; OS sandboxes
Price
Free software; you pay only the model, or use a local one with Ollama
Model
Anthropic, OpenAI, local Ollama and about 20 other providers, with fallback chains
Where you talk to it
30+ channels: Discord, Telegram, Matrix, email, voice, webhooks, command line; web dashboard

Sources: zeroclaw-labs/zeroclaw on GitHub

NanoClaw

nanoco (open source, MIT)Free

Spawn teammates from chat, each with its own bot identity, container and memory; scheduled tasks.

Reliability index75/100
Early release or rolling out: Latest release is a release candidate, v2026.10.1-rc.2. Public for 9 months. 31,088 GitHub stars · 100+ commits in the last 30 days · 8 releases in the last 90 days. Stage 25/40 · Track record 20/30 · Upkeep 30/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyEach agent runs in its own Linux container on your machine or server
Watch the browser liveNot statedNot described in the README
Custom MCP serversNot statedNot named in the README
Asks before risky actionsPartlySafety by isolation: agents see only the folders you mount
Price
Free software, but it uses Claude through your Anthropic account, so the model is paid
Model
Claude, through the Claude Agent SDK
Where you talk to it
Slack (one app per agent), WhatsApp, Telegram, Discord through channel adapters

Sources: nanocoai/nanoclaw on GitHub

nanobot

HKUDS (open source, MIT)Free

Very small Python agent with memory, scheduled automations, long-running goals and multi-agent delegation.

Reliability index68/100
Early release or rolling out: Version 0.x, latest v0.4.0. Public for 9 months. 49,871 GitHub stars · 100+ commits in the last 30 days · 2 releases in the last 90 days. Stage 25/40 · Track record 20/30 · Upkeep 23/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyYour own computer or server; always on only while it runs
Watch the browser liveNot statedWeb search and fetch tools; a live browser view is not described
Custom MCP serversYesMCP integrations listed in the README
Asks before risky actionsNot statedSecurity settings documented; approvals not described in the README
Price
Free software; you pay only the model, or use a free or local one
Model
Any provider you configure, with fallback models
Where you talk to it
Browser WebUI, terminal, Telegram, Discord, Slack, WeChat, Email, Mattermost, Linear

Sources: HKUDS/nanobot on GitHub

ChatGPT dots

OpenAIPaid plan

Works 24/7 on its own cloud computer, voice calls, keeps context across channels, can use your laptop if you allow it.

Reliability index60/100
Early release or rolling out: Rolling out to Pro and Business Premium; beta for Enterprise. Public for 1 month. Last official update 3 November 2026 · no security incident found. Stage 25/40 · Track record 5/30 · Upkeep 30/30
In Italy / EU nowNot statedRolling out in “eligible markets”; OpenAI’s post does not list them
Own computer, works with your PC offYes
Watch the browser liveYesYou can open your dot’s cloud computer at any time to check its work
Custom MCP serversNot statedPlugin ecosystem with 4,000+ apps; MCP not named
Asks before risky actionsYesDefault rules plus Custom rules: allow, ask for approval or block an action
Price
First dot included in ChatGPT Pro or Business Premium at no extra cost; Enterprise, Edu and Healthcare in beta
Model
GPT-6 Astra; GPT-6.1 Terra added on 3 November 2026
Where you talk to it
ChatGPT (desktop, web, mobile), Slack, Microsoft Teams; texting coming
Launched
29 September 2026

Sources: OpenAI: Introducing dots

Grok Bot

SpaceXSIPaid plan

Several bots in parallel that message each other in group chats, sharing one persistent cloud computer with browser, files and terminal.

Reliability index60/100
Early release or rolling out: Launched as a beta according to reviews; the official docs do not say. Public for 2 months (launch date from third-party reviews). Last official update 31 October 2026 · no security incident found. Stage 25/40 · Track record 5/30 · Upkeep 30/30
In Italy / EU nowPartlyNo general region limit stated; tagging @bot on X is not available in the EU
Own computer, works with your PC offYes
Watch the browser liveYesWatch clicks and typing live, take control for a login or CAPTCHA, then hand back
Custom MCP serversYesA Bot can add an MCP server that is not in the catalog, by URL or command
Asks before risky actionsYesSensitive actions can stop for approval; “Ask first” rules
Price
Included in X Premium, Super and Ultra (not Plus or Free), in paid Cursor plans and some SuperGrok plans; free trial on some accounts
Model
Picks the model for each task
Where you talk to it
Apps for macOS, Windows, Linux, iOS and Android; X, Slack, email

Sources: SpaceXSI docs: Grok Bot overview · Grok Bot computer and apps · Grok Bot in the X Subscription · Tag @bot on X

Muse

MetaFree

Talk to it like a contact on WhatsApp, full audit trail, checkout with Link by Stripe and one-time cards.

Reliability index60/100
Generally available: Publicly available in the US and Singapore. Public for 1 month. Last official update 27 October 2026. Stage 40/40 · Track record 5/30 · Upkeep 15/30 Incident: About 6,800 shared chat links were found reachable; Meta paused chat sharing (October 2026)
In Italy / EU nowNoUS since September, Singapore since 27 October 2026
Own computer, works with your PC offYes
Watch the browser liveNot statedA dedicated secure virtual machine with its own browser; a live view is not described by Meta
Custom MCP serversNot statedGmail, Google Calendar, Instagram and shopping partners; MCP not named
Asks before risky actionsYesChecks with you before sending an email or buying; a “Sentinel” approves every request to the internet
Price
Free for most uses, with subscription plans for more (Meta does not state the prices)
Model
Muse Spark
Where you talk to it
Muse app (iOS, Android), muse.ai, WhatsApp
Launched
8 September 2026

Sources: Meta: Introducing Muse · Meta: Muse in Singapore

AnyModelBot

Modelflare (open source, MIT)Free

Presents itself as the open, local-first alternative to Grok Bot, ChatGPT dots and Claude Duet; several bots, each with its own persona, workspace and memory.

Reliability index60/100
Early release or rolling out: Version 0.x, latest v0.1.22. Public for less than a month. 2,722 GitHub stars · 100+ commits in the last 30 days · 21 releases in the last 90 days. Stage 25/40 · Track record 5/30 · Upkeep 30/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyDesktop app for Windows, macOS and Linux; works in the background while your computer is on
Watch the browser liveNot statedNot described in the README
Custom MCP serversNot statedNot named in the README
Asks before risky actionsYesWorks in the background and asks you when a decision is needed
Price
Free app; models through Modelflare (pay per token) or your own provider keys
Model
Any model; each bot has its own default model
Where you talk to it
Desktop app for Windows, macOS and Linux

Sources: modelflare-ai/AnyModelBot on GitHub

Claude Duet

AnthropicPaid plan

Morning brief, checks your apps every few hours, carries over your Claude memory, skills and connected apps.

Reliability index45/100
Preview or waitlist: Research preview with a US waitlist. Public for less than a month. Last official update 21 October 2026 · no security incident found. Stage 10/40 · Track record 5/30 · Upkeep 30/30
In Italy / EU nowNoUS only, waitlist; Team and Enterprise plans not included yet
Own computer, works with your PC offYes
Watch the browser liveNoIt has its own browser in a virtual machine but “can’t yet show you its browser”
Custom MCP serversNot stated6,000+ connectors and plugins from the Claude Marketplace; MCP not named
Asks before risky actionsYesShows what it is about to do and waits before sending, paying, signing or deleting
Price
In Claude Pro, Max and Ultra; no extra price stated, it uses your plan’s usage limits
Model
Claude Opus 5.5
Where you talk to it
Claude Duet app for iOS and Android, claude.ai on the web
Launched
21 October 2026 (research preview)

Sources: Anthropic: Introducing Claude Duet · Claude Duet product page

PicoClaw

Sipeed (open source, MIT)Free

Easiest install: download, double-click, set it up in the browser; schedules and memory.

Reliability index45/100
Early release or rolling out: Version 0.x, last stable v0.2.9 (May 2026), nightly builds since. Public for 9 months. 30,104 GitHub stars · 0 commits in the last 30 days · 0 releases in the last 90 days. Stage 25/40 · Track record 20/30 · Upkeep 0/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offPartlyTiny program for your computer, a server or small boards
Watch the browser liveNot statedWeb UI for settings and chat; a live browser view is not described
Custom MCP serversYesNative MCP support, MCP servers managed from the Web UI
Asks before risky actionsNot statedNot described in the README
Price
Free software; you pay only the model, and simple questions can go to cheaper models
Model
Many providers, with rule-based routing between light and strong models
Where you talk to it
Web UI, Matrix, IRC, WeCom, Discord and others

Sources: sipeed/picoclaw on GitHub

Gemini agent

GooglePaid plan

Coworker agents get their own Workspace account: @agents.company.com email, calendar and Drive.

Reliability index40/100
Announced only: Announced; no availability date, preview only for two industries. Public for less than a month. Last official update 8 October 2026 · no security incident found. Stage 5/40 · Track record 5/30 · Upkeep 30/30
In Italy / EU nowNot statedAvailability and plans not stated; preview only for Financial Services and Legal
Own computer, works with your PC offYes
Watch the browser liveNot statedRuns in the cloud and keeps working after you close your laptop; browser use not described
Custom MCP serversYesConnects to any MCP server inside or outside your company network
Asks before risky actionsPartlyPermissions approved by your security admins; audit trail and real-time spend caps
Price
Not stated; for businesses first
Model
Choice of models, including Gemini and Claude
Where you talk to it
Web, iOS, Android, Windows, Mac, command line, Google Workspace, Microsoft 365, Slack
Launched
Announced 8 October 2026

Sources: Google Cloud: Gemini at Work 2026 · TechCrunch, 8 October 2026

TrustClaw

Composio (open source, MIT)Free

24/7 assistant with scheduled runs, vector memory and Gmail, GitHub, Slack, Notion and Calendar tools.

Reliability index38/100
Preview or waitlist: No releases yet. Public for 6 months. 911 GitHub stars · 1 commits in the last 30 days · 0 releases in the last 90 days. Stage 10/40 · Track record 20/30 · Upkeep 8/30
In Italy / EU nowYesRuns on your own computer or server, anywhere
Own computer, works with your PC offYesYou deploy it to the cloud (for example Vercel); every action runs in a temporary cloud sandbox
Watch the browser liveNot statedNot described in the README
Custom MCP serversNot stated1,000+ Composio tools via OAuth; MCP not named
Asks before risky actionsPartlyTools limited to the accounts you connect; actions run in a sandbox
Price
Free code; needs a Composio API key for the tools and a Claude model
Model
Claude
Where you talk to it
Web dashboard, Telegram

Sources: ComposioHQ/trustclaw on GitHub

“Not stated” means the official page does not say it: it may exist. Prices and availability change often; check the source before you pay.

How the reliability index works

It measures how mature and how well looked after each agent is, from facts anyone can check. It is not a test of answer quality. Out of 100:

  • Stage, up to 40: generally available 40, early release or rolling out 25, preview or waitlist 10, announced only 5.
  • Track record, up to 30: public for 12 months or more 30, 6 to 12 months 20, 3 to 6 months 10, less than 3 months 5.
  • Upkeep, up to 30. Open source: 15 for 50 or more commits in the last 30 days (8 for at least one), plus 15 for 3 or more releases in the last 90 days (8 for at least one). Big labs: 15 for an official update in the last month, plus 15 if we found no security incident.

GitHub numbers measured on 6 November 2026 on each project’s main branch. New products start low on track record by design: being new is a risk, not a flaw.

Free always-on agents

OpenClaw, Hermes Agent, nanobot, ZeroClaw, NanoClaw, PicoClaw, AnyModelBot, TrustClaw are open source and free. You run them on your own computer or on a server, and most answer you on WhatsApp, Telegram, Slack or Discord. They are always on only while that machine is on, so for 24/7 use put them on a small server. The software is free; the model is the main cost, and it can be free too: pick one from our best free LLMs or run one at home. NanoClaw and TrustClaw use Claude, so their model is paid.

Meta’s Muse also has a free tier, but only in the US and Singapore for now.

How to choose

You are in Italy or the EU

Today the safe choices are the free, self-hosted ones. Claude Duet and Muse are not available here yet; ChatGPT dots and Google’s Gemini agent do not say clearly which countries they cover.

You want the most mature free one

OpenClaw and Hermes Agent have the longest track record and the most active upkeep. ZeroClaw asks for approval before medium-risk actions by default.

You want a team of bots

Grok Bot runs several bots that talk to each other in group chats, and you can watch and take over its browser. Free: NanoClaw and AnyModelBot also run several bots.

You live in ChatGPT, Slack or Teams

ChatGPT dots answer in ChatGPT, Slack and Teams and keep the context across them. The first dot is included in Pro and Business Premium.

Share this page:XLinkedInRedditWhatsApp

Press / or Ctrl+K to search from any page.