# OpenClaw alternatives: what to switch to, based on why you are leaving OpenClaw alternatives matched to your reason for leaving: security defaults, upkeep, cost, a hosted assistant, coding or automation. Defaults and maintenance checked October 2026. Author: Ravi Parikh Canonical: https://voyager.so/blog/openclaw-alternatives Published: 2026-10-20T11:33:00-07:00 Updated: 2026-10-20T11:33:00-07:00 ![Table titled Open-source OpenClaw alternatives: defaults and upkeep, read 20 October 2026. Columns: where commands run after a standard install, whether it asks before running a command, stable releases in the last 90 days, and whether it imports an OpenClaw setup. OpenClaw: on the host with the sandbox off, does not ask, 26 releases. Hermes Agent: on the host, asks only about commands it flags as dangerous and unattended jobs refuse them, 19, imports. NanoClaw: in a Docker container per session, does not ask inside the container, 5, imports. Moltis: in a container if Docker, Podman or Apple Container is installed, otherwise on the host; asks on the host about commands not on its allowlist, 16, imports. ZeroClaw: on the host with an OS sandbox where available, asks about medium-risk commands unless you pick the recommended Balanced preset, 3, imports memory only. IronClaw: on the host, with WebAssembly isolation only for extension tools, auto-approve on, 6, no importer found. nanobot: on the host with the sandbox off, no general approval step, 2, no importer found. PicoClaw: on the host, limited to its workspace by path checks, no general approval step, 0, imports.](https://cdn.moda.app/voyager/blog/openclaw-alternatives/hero-9050766a03c5a506.webp) **The right OpenClaw alternative depends on what made you look, and for several common complaints the fix is a setting in OpenClaw rather than a new project.** If you want the agent isolated from your computer, NanoClaw runs every chat session in its own container by default. If running it wore you down, Hermes Agent is the closest like-for-like replacement, with a desktop app that sets itself up and tested releases that install only after passing a check. If you want nothing of your own to run, OpenAI's dots and Claude Cowork are vendor-run versions of the idea. If you mostly wanted to text a coding agent, or to automate the same steps every time, there are better tools than a general assistant. | If you are leaving because | Try | The catch | | ------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- | | You want the agent walled off from your computer | [NanoClaw](#if-you-want-the-agent-walled-off-from-your-computer), or OpenClaw with `ask` mode and its sandbox on | Inside NanoClaw's container the agent runs without prompts, with open outbound network access by default | | Upkeep, but you like OpenClaw | [A managed OpenClaw host](#if-you-left-because-of-setup-and-upkeep) | You still choose the skills and grant the account access | | Upkeep, and you are done with OpenClaw | [Hermes Agent or Hermes Cloud](#if-you-left-because-of-setup-and-upkeep) | Its importer leaves scheduled jobs, plugins and multi-agent setups for you to rebuild | | You want nothing to run yourself | [OpenAI dots or Claude Cowork](#if-you-want-someone-else-to-run-all-of-it) | Dots need ChatGPT Pro ($100 a month and up) or Business Premium; Cowork has no chat-app channel | | The bill | [OpenClaw on a ChatGPT plan, with a longer heartbeat](#if-you-left-because-of-the-bill) | Subscription routes have usage caps, and Claude routes can be billed as extra usage | | You mainly code | [Claude Code channels or Codex cloud](#if-what-you-wanted-was-a-coding-agent) | Built for repositories, not your inbox | | The same steps every time | [n8n](#if-you-want-the-same-steps-every-time) | You build each workflow yourself | | A Raspberry Pi or home server | [ZeroClaw](#if-you-want-it-local-or-on-small-hardware) | A smaller community, and its sandbox depends on the build and OS | Everything here comes from each project's documentation, source code and release notes, and each vendor's help centre and pricing pages, read on 20 October 2026, with release figures from the GitHub API that day. No agent was installed for this guide, so it covers what each is documented and coded to do, not how reliably it does it. Prices are in US dollars unless stated, as of 20 October 2026. Voyager, the desktop app our team builds, [runs Claude Code and Codex on your existing sign-ins](/blog/why-we-forked-t3-code) for desktop creative work in apps such as Photoshop and Blender. ## First, check whether a setting fixes it OpenClaw's defaults favour getting started over caution. Before you migrate your memory, skills and chat connections somewhere else, check whether one of these is the real problem. | If the problem is | Change this in OpenClaw | When that is not enough | | -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- | | It runs commands without asking | Set the command mode to `ask` (prompts for anything not on your allowlist) or `auto` (a reviewer model allows, refuses or asks) ([permission modes](https://docs.openclaw.ai/tools/permission-modes)). Scheduled jobs then refuse a command that needs approval when nobody can answer ([exec approvals](https://docs.openclaw.ai/tools/exec-approvals)) | You want a boundary, not prompts | | It can reach everything on the machine | Turn on the tool sandbox (Docker, Podman, SSH or OpenShell backends), or run the whole Gateway in Docker ([sandboxing](https://docs.openclaw.ai/gateway/sandboxing), [Docker](https://docs.openclaw.ai/install/docker)) | Plugins still run inside the Gateway with its full authority | | Model bills you did not expect | Lengthen or stop the heartbeat, a check-in that wakes the agent every 30 minutes by default, or hourly when Claude is connected with a subscription sign-in; `heartbeat.every: "0m"` turns it off. Watch `/usage cost` ([heartbeat](https://docs.openclaw.ai/gateway/heartbeat), [token use](https://docs.openclaw.ai/reference/token-use)) | The rest of the bill is the model you chose and how much you ask of it | | Updates keep breaking it | Move to the extended-stable channel, which trails by a supported month and never updates automatically, and back up before each upgrade with `openclaw backup create --verify` ([release channels](https://docs.openclaw.ai/install/development-channels), [backup](https://docs.openclaw.ai/cli/backup)) | You want someone else to do upgrades | | Strangers could reach it | Leave the Gateway on its default loopback address with its generated token, reach it over Tailscale or an authenticated tunnel, and approve each chat pairing ([security](https://docs.openclaw.ai/gateway/security)) | You need people who do not trust each other on one agent: OpenClaw says that needs separate installations ([teams](https://docs.openclaw.ai/start/teams)) | ## The open-source alternatives, side by side Every project below is a self-hosted agent you message from chat apps, with memory and scheduled jobs, and each is free software you pair with a model you pay for. The rows that matter most are the defaults: what happens after a standard install, before you change any settings. | | OpenClaw | Hermes Agent | NanoClaw | Moltis | ZeroClaw | IronClaw | nanobot | PicoClaw | | ------------------------------------------- | ----------------------------- | -------------------------------------------------------------------------- | -------------------------------------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------------------- | ------------------------------ | --------------------------------------------- | | Maintainer, licence | OpenClaw Foundation, MIT | Nous Research, MIT | NanoCo, MIT | moltis-org, MIT | Community, MIT or Apache 2.0 | NEAR AI, MIT or Apache 2.0 | HKU Data Intelligence Lab, MIT | Sipeed, MIT | | Latest stable release as of 20 October 2026 | v2026.10.5, 18 Oct | v0.22.2, 15 Oct | v2026.10.0, 9 Oct | 20261019.01, 19 Oct | v0.8.6, 18 Oct | v1.4.1, 29 Sep | v0.3.5, 15 Sep | v0.3.1, 3 Jul | | Where commands run by default | Host, sandbox off | Host | Docker container per session | Container if Docker, Podman or Apple Container is installed; otherwise host | Host, with an OS sandbox when one is available | Host; WebAssembly only for extension tools | Host, sandbox off | Host, with application-level workspace checks | | Asks before running a command by default | No | Only commands matching its dangerous patterns; unattended jobs refuse them | No, inside the container | Only on the host, for commands not on its allowlist | Medium-risk commands; the recommended Balanced preset stops asking | No; auto-approve is on | No general approval step | No general approval step | | API keys | Readable by the agent process | Readable by the agent process | Kept outside the container by a credential gateway | Encrypted at rest, passed to commands at run time | Encrypted store, read by the agent process | Hidden from WebAssembly tools; the host shell gets a scrubbed environment | Readable by the agent process | Readable by the agent process | | Imports an OpenClaw setup | n/a | `hermes claw migrate` | `/migrate-from-openclaw` skill | `moltis import` | `zeroclaw migrate openclaw` (memory) | None found | None found | `picoclaw migrate --from openclaw` | Sources, all read 20 October 2026, at each project's latest stable release where the file exists there: OpenClaw's [exec tool](https://docs.openclaw.ai/tools/exec) and [security policy](https://github.com/openclaw/openclaw/blob/v2026.10.5/SECURITY.md); Hermes's [security guide](https://hermes-agent.nousresearch.com/docs/user-guide/security) and [SECURITY.md](https://github.com/NousResearch/hermes-agent/blob/v0.22.2/SECURITY.md); NanoClaw's [security model](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/docs/SECURITY.md); Moltis's [sandbox](https://github.com/moltis-org/moltis/blob/20261019.01/docs/src/sandbox.md) and [approval code](https://github.com/moltis-org/moltis/blob/20261019.01/crates/tools/src/exec.rs); ZeroClaw's [sandboxing](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/docs/book/src/security/sandboxing.md), [autonomy](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/docs/book/src/security/autonomy.md) and [risk presets](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/crates/zeroclaw-config/src/presets.rs); IronClaw's [deployment profile](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/app/ironclaw_composition/src/deployment.rs), [runtime policy](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/kernel/ironclaw_runtime_policy/src/resolver.rs) and [auto-approve default](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/kernel/ironclaw_approvals/src/auto_approve.rs); nanobot's [shell settings](https://github.com/HKUDS/nanobot/blob/v0.3.5/nanobot/agent/tools/shell.py); PicoClaw's [defaults](https://github.com/sipeed/picoclaw/blob/v0.3.1/pkg/config/defaults.go) and [shell tool](https://github.com/sipeed/picoclaw/blob/v0.3.1/pkg/tools/shell.go); releases from the GitHub API. After a standard install, five of these seven alternatives run the agent's commands directly on your computer, as OpenClaw does, with different restrictions around them; only NanoClaw always uses a container, and Moltis uses one when a container runtime is installed. Prompts before commands are rarer still. On command execution, most of the protections these projects advertise are options you turn on, the same as OpenClaw's. Filesystem and process isolation limit access to your computer; connected-account access also depends on credential scope and action approvals. NanoClaw's credential gateway supports different policies per agent group, with available limits and approval options depending on the gateway. NanoClaw routes held requests to authorized approvers and denies them on errors ([security model](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/docs/SECURITY.md#4-credential-isolation-credential-gateway)). Use those controls alongside separate accounts or narrowly scoped tokens. On chat apps: all eight cover Telegram, and all but IronClaw (Telegram and Slack only in v1.4.1) cover WhatsApp ([IronClaw channels](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/docs/channels/overview.mdx)). iMessage is the outlier: OpenClaw, Hermes, NanoClaw and ZeroClaw (experimentally) support it, and each documented route needs a Mac signed in to Messages or a hosted relay service ([OpenClaw iMessage](https://docs.openclaw.ai/channels/imessage), [Hermes Photon](https://hermes-agent.nousresearch.com/docs/user-guide/messaging/photon), [ZeroClaw channels](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/docs/book/src/channels/chat-others.md)). ## If you want the agent walled off from your computer OpenClaw describes itself as infrastructure "for trusted operators" ([security policy](https://github.com/openclaw/openclaw/blob/v2026.10.5/SECURITY.md)): out of the box its command tool runs on the host with no approval prompt and no sandbox ([exec tool](https://docs.openclaw.ai/tools/exec)). Its history adds to the worry: more than 21,000 instances were visible on the internet in January, most still behind a token, according to [Censys](https://censys.com/blog/openclaw-in-the-wild-mapping-the-public-exposure-of-a-viral-ai-assistant/), and its skill registry was used to distribute malware in February, according to [VirusTotal](https://blog.virustotal.com/2026/02/from-automation-to-infection-how.html). Four projects pitch themselves as the fix, and they mean different things by it. **NanoClaw: a container per session.** Its README argues that OpenClaw's protections are application-level, allowlists and pairing codes, rather than operating-system isolation ([README](https://github.com/nanocoai/nanoclaw)). In NanoClaw each session runs in a Docker container as a non-root user and sees only folders that are explicitly mounted; the project directory is never mounted, extra folders are blocked unless an allowlist outside the agent's reach permits them, and API keys stay outside the container, added to outgoing requests by a credential gateway on the host ([security model](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/docs/SECURITY.md)). The container is the main boundary, and three things sit inside it. The agent runs Claude's Agent SDK with permission prompts bypassed ([agent settings](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/container/agent-runner/src/providers/claude-config.ts)), so anything you mount read-write is the agent's to change. Outbound network access is open until you turn on its lockdown setting. And sessions in the same agent group share that group's files and memory. On 13 October the project published nine advisories, six rated high, covering setup, permissions and the host's handling of agent-writable folders, all addressed by v2026.10.0 ([advisories](https://github.com/nanocoai/nanoclaw/security/advisories)). It needs Docker, Node 22 and pnpm, defaults to Claude (an Anthropic API key or subscription), and adds Codex or OpenCode for other models through skills. Customising it means changing its code, usually by asking Claude Code to do it. **Moltis: a container when one is available.** This single Rust binary picks the strongest sandbox it finds, in order Apple Container, Podman, Docker, then a "restricted host" mode with "no filesystem isolation", then the host itself, with a startup warning ([sandbox](https://github.com/moltis-org/moltis/blob/20261019.01/docs/src/sandbox.md)). In a real container it runs commands without asking; on the host it asks about any command not on its allowlist ([approval code](https://github.com/moltis-org/moltis/blob/20261019.01/crates/tools/src/exec.rs)). Its container traffic goes through a "trusted network" proxy that allows listed destinations by default ([config](https://github.com/moltis-org/moltis/blob/20261019.01/crates/config/src/schema/tools.rs)), but the docs warn that some tools do not use the proxy, so treat it as filtering rather than a sealed network ([trusted network](https://docs.moltis.org/trusted-network.html)). Two more defaults to know: every session shares one sandbox home folder unless you set `home_persistence = "session"`, and if Docker loses its connection mid-session, Moltis switches to the restricted-host mode and stays there until you restart it ([sandbox](https://github.com/moltis-org/moltis/blob/20261019.01/docs/src/sandbox.md)). Install Docker or Podman first if isolation is why you are here, and watch for that switch. It published two advisories in September, one critical, both fixed ([advisories](https://github.com/moltis-org/moltis/security/advisories)). **IronClaw: isolation for tools, not for the shell.** NEAR AI's Rust agent runs extension tools in WebAssembly sandboxes that can reach only the hosts and credentials declared for them, with secrets injected at the boundary so tool code never sees them ([sandboxed tools](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/docs/capabilities/sandboxed-tools.mdx)). Its built-in shell is different. In v1.4.1's default local profile, commands run as an ordinary process on your computer ([deployment profile](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/app/ironclaw_composition/src/deployment.rs), [runtime policy](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/kernel/ironclaw_runtime_policy/src/resolver.rs)), and a global auto-approve setting is on by default, with a code comment explaining it is on "so new users aren't prompted per tool call" ([auto-approve](https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/crates/kernel/ironclaw_approvals/src/auto_approve.rs)). NEAR AI also hosts it: a Starter plan shown as free with $5 of credits for one agent, and Basic at $20 a month with $20 of credits for up to two ([IronClaw](https://www.ironclaw.com/)). **ZeroClaw: a policy engine, with OS sandboxes where the build supports them.** Its built-in default is "supervised": medium-risk commands need approval, high-risk ones are blocked unless you allowlist them explicitly, and file access is limited to its workspace ([autonomy](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/docs/book/src/security/autonomy.md)). The setup wizard's recommended Balanced preset keeps the high-risk blocking but stops asking about medium-risk commands; the Locked Down preset keeps asking ([presets](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/crates/zeroclaw-config/src/presets.rs)). It runs on the host and adds an OS sandbox when it can: on standard builds, Firejail on Linux and Seatbelt on macOS; Landlock and Bubblewrap need a custom build; Windows has none; and if detection fails it runs unsandboxed ([sandboxing](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/docs/book/src/security/sandboxing.md)). Hermes Agent is not in this group: it runs on the host and asks only about commands matching its dangerous patterns, and its own policy says "The only security boundary against an adversarial LLM is the operating system" ([SECURITY.md](https://github.com/NousResearch/hermes-agent/blob/v0.22.2/SECURITY.md)). Its useful default is narrower: scheduled and unattended jobs refuse a command that would need approval ([security guide](https://hermes-agent.nousresearch.com/docs/user-guide/security)), which OpenClaw also does once you switch on `ask` mode. Choose NanoClaw if you want isolation that does not depend on a setting, and turn on its network lockdown. Choose Moltis if you want the same idea in a single binary, with a filtering proxy by default, and will install a container runtime. Otherwise, compare OpenClaw with `ask` mode and its sandbox on against any alternative with the same settings: on command execution, most of these projects start closer to OpenClaw than their pitches suggest. ## If you left because of setup and upkeep OpenClaw's installer is quick, and the docs promise a first chat in about five minutes ([getting started](https://docs.openclaw.ai/start/getting-started)). The work comes after: a machine that stays on, a token or QR code per chat app, sign-ins for each account a skill touches, and frequent updates, with 26 non-prerelease releases across its release channels in the 90 days to 20 October ([releases](https://github.com/openclaw/openclaw/releases)). You can shrink that work three ways. **Run Hermes Agent instead.** It does the same job, one gateway process connected to more than twenty chat platforms, with scheduled jobs, memory and self-written skills, and its 13 October release targeted setup and updates. The desktop app, a signed package for Windows and Apple Silicon Macs that carries its own Python and Node, opens on a guided setup chat running on a free trial model, so you need no API key to start. Installs now follow tested SemVer releases, and an update builds a new environment and switches only once it passes a check ([v0.22.0 notes](https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.0)). Its always-loaded memory is two files with configurable default caps of 2,200 and 1,375 characters, which you can read at a glance; older conversations stay searchable, and plugins can add more ([memory](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory)). None of that guarantees fewer breakages, and people have moved in both directions; [Hermes vs OpenClaw](/blog/hermes-vs-openclaw) collects those accounts. The switch costs work: what its importer leaves behind is in the [migration table](#moving-your-setup). **Let someone run OpenClaw for you.** The OpenClaw Foundation sells no hosted service ([FAQ](https://docs.openclaw.ai/help/faq/what-is-openclaw)), but hosting companies do, and they differ in what they take on. Hostinger's Managed OpenClaw says it handles "all security, updates, and backups" and sells model credits in its dashboard; it shows $4.99 a month, paid upfront for the term, renewing at $11.99 a month for two years ([Hostinger](https://www.hostinger.com/managed-openclaw)). DigitalOcean's one-click Droplet is a starting point rather than a managed service: from $12 a month, it ships with a gateway token, firewall rate limits, a non-root user, Docker sandboxing and DM pairing, but the server and its updates are yours, its image listed OpenClaw 2026.9.3 on 20 October, and its own troubleshooting FAQ suggests switching approval prompts off ([DigitalOcean](https://docs.digitalocean.com/products/marketplace/catalog/openclaw/)). Either way you still choose the skills and grant the account access. **Let Nous run Hermes for you.** Hermes Cloud runs a Hermes agent on a dedicated cloud instance with its own dashboard, reachable from Telegram, Discord, Slack and email. It bills daily from prepaid Nous credit for compute metered by the second and for storage (including while stopped), plus model and tool use; deploying needs $2 of credit or a Nous subscription, and instance prices are shown only on the deploy screen ([Hermes Cloud](https://portal.nousresearch.com/cloud)). Choose Hermes if you want guided setup and a more cautious update path, and can rebuild your scheduled jobs. Choose Hostinger's managed plan if you like OpenClaw and dislike being its system administrator. ## If you left because of the bill OpenClaw is free; you pay for a machine and a model, and the agent spends tokens when you are not talking to it. What each option charges for, as of 20 October 2026: | Option | Fixed monthly cost | Model use | | ---------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | OpenClaw or an open-source alternative on your own machine | None for the software; a small server from about $6 a month ([OpenClaw's DigitalOcean guide](https://docs.openclaw.ai/install/digitalocean)) | Per token with an API key, a subscription route, or free on local hardware | | Hostinger Managed OpenClaw | $4.99 a month paid upfront, then $11.99 ([Hostinger](https://www.hostinger.com/managed-openclaw)) | Credits bought in its dashboard, or your ChatGPT account | | Hermes Cloud | Daily instance and storage charges from prepaid credit ([Hermes Cloud](https://portal.nousresearch.com/cloud)) | Billed from the same credit | | Claude Cowork | Claude Pro $20, Max $100 or $200 ([pricing](https://claude.com/pricing)) | Included up to the plan's five-hour and weekly limits | | OpenAI dots | ChatGPT Pro $100, $200 or $500 ([Pro tiers](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers)) | An included allowance for deeper work | | Manus | Free (300 daily credits, one task at a time, 2 scheduled automations) or Pro from $20 ([pricing](https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus)) | Monthly credits, 4,000 on the $20 Pro plan | | n8n | Free self-hosted, or Cloud from €20 a month billed annually for 2,500 workflow runs ([pricing](https://n8n.io/pricing/)) | Your provider account and API charges, or prepaid Gateway credits on Cloud Starter/Pro for supported nodes ([billing options](https://support.n8n.io/article/can-i-use-my-own-ai-api-key-on-n-8-n-cloud)) | This compares how each option bills, not what a given workload costs, so treat it as the shape of each bill rather than a ranking. If you stay with OpenClaw, the heartbeat in the first table, the model and how long a chat runs before you reset it are your controls, plus two routes: * **A subscription you already pay for.** OpenClaw can run turns through OpenAI's Codex app-server on a ChatGPT plan, and its docs say OpenAI supports that ([OpenAI provider](https://docs.openclaw.ai/providers/openai)). It can also drive the Claude Code program signed in with a Claude plan, though Anthropic reserves the right to bill third-party tools as extra usage instead of plan limits ([Anthropic provider](https://docs.openclaw.ai/providers/anthropic), [Anthropic's login article](https://support.claude.com/en/articles/13189465-log-in-to-your-claude-account)). ZeroClaw, Moltis and PicoClaw also document ChatGPT sign-in; Hermes supports it but says plan-quota accounting is not documented ([Hermes providers](https://hermes-agent.nousresearch.com/docs/integrations/providers)). * **A cheaper model, carefully.** OpenClaw's own FAQ warns: "Smaller/quantized models raise prompt-injection risk" ([hosting FAQ](https://docs.openclaw.ai/help/faq-first-run/providers-and-hosting)). A cheap model is lower risk on jobs with no tools that can send, buy or delete, though anything it can read it can still leak. ## If you want someone else to run all of it These run on the vendor's computers and keep working when you close the app. They differ from OpenClaw in three ways: you choose models only within the vendor's family (Hermes Cloud excepted), email and calendar come as ready-made connections instead of skills you install, and only some let you text them from the apps you already use. | | OpenAI dots | Claude Cowork | Manus | Hermes Cloud | | ------------------------------------ | ---------------------------------------------------------------- | ----------------------------------------------------------- | ---------------------------------------------------------------- | ---------------------------------------------- | | Plan needed | ChatGPT Pro or Business Premium | Claude Pro or Max | Free tier; Pro from $20 a month | Prepaid Nous credit | | Where you message it | ChatGPT apps, Slack, Teams; texting is a limited US beta for Pro | Claude web, desktop and phone apps | Manus apps and Telegram | Telegram, Discord, Slack, email, its dashboard | | Keeps working with your computer off | Yes, on its own cloud computer | Yes, for new Pro and Max tasks | Yes, in a cloud virtual machine | Yes | | Can use your own computer | Optional; the ChatGPT desktop app must be open on it | Folders, browser and apps, with the Claude desktop app open | Through the Manus desktop app, approving each folder and command | No | | Models | GPT-6 Astra | Claude models | Manus's own agents | Your choice | Sources, read 20 October 2026: OpenAI's [introducing dots](https://openai.com/index/introducing-dots/) and [getting started with your dot](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot); Anthropic's [Claude pricing](https://claude.com/pricing) and [Cowork on web, desktop and mobile](https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile); Manus's [pricing](https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus), [Telegram launch](https://manus.im/blog/manus-agents-telegram) and [desktop app](https://www.manus.im/blog/manus-my-computer-desktop); [Hermes Cloud](https://portal.nousresearch.com/cloud). **OpenAI dots** are the closest commercial match to OpenClaw's pitch. Launched on 29 September, a dot is "always-on", has its own cloud computer and browser, connects through plugins to "over 4,000 apps", learns your preferences, and works in the background ([announcement](https://openai.com/index/introducing-dots/)). Pro users outside the European Economic Area, Switzerland and the UK, and Business Premium users wherever ChatGPT is supported, get the first dot as part of their plan ([help article](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot)). Its defaults are the reverse of OpenClaw's: background research uses read-only tools, an automatic reviewer checks actions that could affect your accounts and decides whether they proceed, need your approval or need you to do them, and Custom Rules let you set, per kind of action, whether it acts, acts only when you explicitly asked, asks first, or hands off to you. Two things to know before relying on it: disconnecting an app does not delete what the dot already learned from it (only resetting the dot does), and texting it is a limited beta for US Pro users through a third-party provider. ![OpenAI Help Center article Getting started with your dot, section Which plans include dots. It says dots are rolling out to Pro users in markets excluding the European Economic Area, Switzerland and the UK, are available to Business Premium users in all supported ChatGPT regions, and that Enterprise users can try a beta when an admin enables it. It adds that the first dot is included in Pro or Business Premium at no extra cost, with an allowance for deeper work.](https://cdn.moda.app/voyager/blog/openclaw-alternatives/fig-dots-cb2cb664d1b9bba0.webp) Who can get a dot, and what it costs. From OpenAI's Help Center, 20 October 2026. **Claude Cowork** turns files, inbox and documents into finished work: spreadsheets, decks, reports ([get started](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork)). Anthropic is folding it into ordinary Claude conversations, so some accounts no longer show a separate Cowork mode ([Cowork and chat are one Claude](https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude)). Gmail, Google Calendar and Drive connect as connectors, and sending, replying and forwarding ask for your approval by default ([Google Workspace connectors](https://support.claude.com/en/articles/10166901-use-google-workspace-connectors)). Since 6 October, new Pro and Max tasks run in Anthropic's cloud and keep going with your laptop shut, and scheduled tasks run "with no device online"; anything on your own computer still needs the Claude desktop app open there ([what changed on 6 October](https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile)). What it lacks is a chat-app channel: you talk to it in Claude's own apps, and Dispatch, the feature for sending tasks from your phone to your desktop, "isn't available to new users" ([Dispatch](https://support.claude.com/en/articles/13947068-assign-tasks-from-anywhere-in-claude-cowork)). **Manus** runs each task in a cloud virtual machine and added Telegram on 16 February for every plan: you can send voice notes, photos and files, start multi-step research or reports, and get recurring summaries in the chat, and the agent sees only messages sent to it ([Telegram launch](https://manus.im/blog/manus-agents-telegram)). The free plan gives 300 credits a day, one task at a time and 2 scheduled automations; Pro starts at $20 a month for 4,000 credits and 20 automations ([pricing](https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus)). Credits run out faster on long tasks, so try the free plan on your real jobs first. ChatGPT agent "is no longer available"; OpenAI points people to ChatGPT Work instead ([help article](https://help.openai.com/en/articles/11752874-chatgpt-agent)). Perplexity's Computer can be messaged in Slack and runs scheduled work while you are offline, on Perplexity's paid plans with a paid Slack workspace ([Perplexity](https://www.perplexity.ai/try-computer/slack)). Choose dots if you want the OpenClaw experience with OpenAI running it and have Pro or Business Premium. Choose Cowork if your work ends in documents and you already pay for Claude. Choose Manus if you want a Telegram assistant on a cheaper plan. Choose Hermes Cloud if you want hosting but still want to pick the model. ## If what you wanted was a coding agent If your OpenClaw setup mainly relays requests to a coding session, OpenClaw's own FAQ points elsewhere: "Use Claude Code or Codex for the fastest direct coding loop inside a repo" ([What is OpenClaw?](https://docs.openclaw.ai/help/faq/what-is-openclaw)). Both now reach your phone directly: * **Claude Code channels** connect Telegram, Discord or iMessage to a running Claude Code session; they are a research preview, and "Events only arrive while the session is open" ([channels](https://code.claude.com/docs/en/channels)). **Remote Control** drives a session on your computer from the Claude app, with the computer left on ([Remote Control](https://code.claude.com/docs/en/remote-control)). **Cloud scheduled tasks** run repository chores on Anthropic's servers without your laptop ([scheduled tasks](https://code.claude.com/docs/en/routines)). * **Codex cloud** runs coding tasks on OpenAI's computers that you start and continue from web, phone or desktop, on Plus and above ([Codex cloud](https://learn.chatgpt.com/docs/cloud)). In a ChatGPT workspace whose admin has set it up with a shared environment, @ChatGPT in Slack can hand repository work to it ([Slack](https://learn.chatgpt.com/docs/third-party/slack)). ## If you want the same steps every time An agent decides what to do next. Sometimes you want a fixed sequence instead: when an email from a client arrives, summarise it and post it to Slack. That is workflow automation, and n8n is the self-hostable option built for it. The workflow's steps are fixed even when an AI step inside it is not: its AI Agent node gives a workflow a model plus tools, and any tool can be set to wait for human approval through Slack, Telegram, email or other channels before it runs ([AI Agent node](https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent.md), [human review](https://docs.n8n.io/build/integrate-ai/ai-examples/human-in-the-loop-for-tools.md)). Workflows can start from a [Telegram](https://docs.n8n.io/integrations/builtin/trigger-nodes/n8n-nodes-base.telegramtrigger/) message, a chat widget, or [WhatsApp Business Cloud](https://docs.n8n.io/integrations/builtin/trigger-nodes/n8n-nodes-base.whatsapptrigger/), not a personal WhatsApp account. If you would rather not host it, n8n Cloud starts at €20 a month billed annually for 2,500 workflow runs ([pricing](https://n8n.io/pricing/)). The self-hosted Community edition is free under the Sustainable Use License, which covers internal business, personal and non-commercial use; it is not an open-source licence, and charging others for access to n8n you host falls outside it ([licence](https://github.com/n8n-io/n8n/blob/master/LICENSE.md)). Model use is separate from workflow runs: bring your own provider API key on any plan, or use prepaid Gateway credits on Cloud Starter or Pro for supported nodes, with no provider account or API key required ([billing options](https://support.n8n.io/article/can-i-use-my-own-ai-api-key-on-n-8-n-cloud)). Gateway credits pay for model and tool usage in workflows, not n8n Assistant, which has separate credits ([Gateway credits](https://docs.n8n.io/deploy/use-n8n-cloud/gateway-credits)). Choose n8n for jobs you could draw as a flowchart, and keep an agent for open-ended requests. ## If you want it local or on small hardware A local model removes that model's API charges and keeps its requests on your hardware, as long as you configure no hosted fallback. Messages to a Telegram or WhatsApp bot still pass through Telegram or Meta, and web search and other tools still reach their services. Your options: * **OpenClaw or Hermes with a local model.** OpenClaw's smallest curated local recipe is a 4B model with an 8 GiB host-memory floor; setup picks a recipe for your hardware and checks that a real tool call works before switching ([local model catalog](https://docs.openclaw.ai/plugins/llama-cpp)). Hermes refuses to start below 64,000 tokens of context, which Ollama's default does not meet on most machines ([providers](https://hermes-agent.nousresearch.com/docs/integrations/providers)). Small models fail tool calls more often, and OpenClaw's FAQ warns that they raise prompt-injection risk, so test real tasks first. * **ZeroClaw** for a Raspberry Pi or a home server: one binary, Ollama support, and GPIO, I2C, SPI and USB access on boards such as the Pi and ESP32 ([README](https://github.com/zeroclaw-labs/zeroclaw)). * **PicoClaw** advertises a footprint under 10 MB on $10 hardware, while noting recent builds use 10 to 20 MB. Read its own warning first: "Do not deploy to production before v1.0" ([README](https://github.com/sipeed/picoclaw)). Its last release was on 3 July, and a channel with an empty sender list accepts messages from everyone ([channel code](https://github.com/sipeed/picoclaw/blob/v0.3.1/pkg/channels/base.go)). ![PicoClaw README security notice on GitHub: no official crypto tokens, picoclaw.io is the only official website, beware of third-party domains, PicoClaw is in early rapid development and may have unresolved security issues so do not deploy to production before v1.0, and recent builds may use 10 to 20 MB of RAM.](https://cdn.moda.app/voyager/blog/openclaw-alternatives/fig-picoclaw-7e87fc4d16acac3a.webp) PicoClaw's own security notice. From the project's README on GitHub, 20 October 2026. Choose ZeroClaw for small always-on hardware you control. Use PicoClaw for experiments, not for an assistant holding your email. ## Projects to skip for now Three projects that circulate as OpenClaw alternatives have had no commit on their main branch for more than 160 days as of 20 October 2026: [TinyAGI](https://github.com/TinyAGI/tinyagi), formerly TinyClaw (last commit 30 March), [OpenFang](https://github.com/RightNow-AI/openfang) (12 May) and smallnest's [GoClaw](https://github.com/smallnest/goclaw) (27 March). Waren Gonzaga's separate [Tiny Claw](https://github.com/warengonzaga/tinyclaw) was last updated on 25 February. Vellum's hosted assistant closes on 2 November 2026, and its open-source assistant is no longer actively maintained ([shutdown notice](https://www.vellum.ai/vellum-assistant-platform-shutdown)). An agent that holds your email needs a project that is still shipping fixes. ## Moving your setup Five alternatives ship an importer, and they bring over different amounts: | Importer | What comes across | What you redo by hand | | --------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | | Hermes: `hermes claw migrate` | Persona, memories, skills, model, MCP servers and chat-app tokens and allowlists; dry run and backup first; secrets only with `--migrate-secrets` | Scheduled jobs, plugins, webhooks, multi-agent setups and channel bindings (archived for reference); WhatsApp pairing; reset timers | | NanoClaw: `/migrate-from-openclaw` skill, guided by Claude Code | Identity, channels and credentials, workspace and skills, and scheduled tasks, each with your confirmation; it never copies silently | WhatsApp pairing; settings with no NanoClaw equivalent | | Moltis: `moltis import` | Provider keys, memory, skills, sessions, identity, Telegram and Discord settings, MCP servers and workspace files; your OpenClaw install is never modified | Other chat channels | | ZeroClaw: `zeroclaw migrate openclaw` | Memory | Everything else | | PicoClaw: `picoclaw migrate --from openclaw` | Channels, agents and workspace files | Skills, scheduled jobs, webhooks, memory settings, and API keys and OAuth tokens (deliberately skipped) | Sources, read 20 October 2026: [Hermes migration guide](https://hermes-agent.nousresearch.com/docs/guides/migrate-from-openclaw), [NanoClaw migration skill](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/.claude/skills/migrate-from-openclaw/SKILL.md), [Moltis OpenClaw import](https://github.com/moltis-org/moltis/blob/20261019.01/docs/src/openclaw-import.md), [ZeroClaw migration code](https://github.com/zeroclaw-labs/zeroclaw/blob/v0.8.6/crates/zeroclaw-runtime/src/migration.rs), [PicoClaw migration code](https://github.com/sipeed/picoclaw/blob/v0.3.1/pkg/migrate/sources/openclaw/openclaw_config.go). Whichever you use, back up first with `openclaw backup create --verify`, and expect to sign in to websites again. Run the new agent alongside the old one for a week on read-only jobs, such as briefings delivered only to you, before letting it send or change anything; two agents with the same accounts can each send the same email. OpenClaw's memory is plain Markdown (`USER.md`, `MEMORY.md` and daily notes), so a hosted assistant can be seeded by pasting in the parts you want it to know ([memory](https://docs.openclaw.ai/concepts/memory)). ## Frequently asked questions **What is the safest OpenClaw alternative?** If the harm you want to prevent is the agent reading or changing files on your computer, NanoClaw, which always runs the agent in a container that sees only folders you mount; its outbound network access stays open until you lock it down ([security model](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/docs/SECURITY.md)). Filesystem and process isolation alone do not define what an agent may do through connected accounts. NanoClaw also supports per-group credential policies and approval flows, depending on its gateway; review these alongside account and token permissions ([credential controls](https://github.com/nanocoai/nanoclaw/blob/v2026.10.0/docs/SECURITY.md#4-credential-isolation-credential-gateway)). Among hosted assistants, OpenAI's dots and Claude Cowork check or ask before consequential actions by default ([dots](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot), [Cowork](https://support.claude.com/en/articles/13364135-use-claude-cowork-safely)). **NanoClaw vs OpenClaw: what do you give up?** You gain a container per session and a codebase small enough to read ([README](https://github.com/nanocoai/nanoclaw)). You give up OpenClaw's companion apps for phones and laptops, its team mode, its larger skill registry and some of its chat channels ([OpenClaw](https://github.com/openclaw/openclaw)), and NanoClaw's importer asks you to confirm each piece it moves. NanoClaw is also Claude-first; other models come through add-on skills. **Does OpenAI have an OpenClaw alternative?** Yes: dots, launched on 29 September 2026, always-on agents with their own cloud computer that you message in ChatGPT, Slack or Teams, included with ChatGPT Pro and Business Premium ([announcement](https://openai.com/index/introducing-dots/)). Separately, OpenAI donates to the OpenClaw Foundation and employs OpenClaw's creator, but does not own or direct the project ([FAQ](https://docs.openclaw.ai/help/faq/what-is-openclaw)). **Is OpenClaw still maintained?** Very actively: version 2026.10.5 shipped on 18 October 2026, with a newer beta the next day ([releases](https://github.com/openclaw/openclaw/releases)). It is run by the OpenClaw Foundation, a nonprofit announced in July ([Foundation](https://openclaw.ai/blog/introducing-openclaw-foundation)). **Is there a free OpenClaw alternative?** All the open-source agents here are free, as is OpenClaw itself; you pay for the model and the machine ([OpenClaw FAQ](https://docs.openclaw.ai/help/faq/what-is-openclaw)). A local model removes the model bill but needs capable hardware. Among hosted options, Manus has a free tier with 2 scheduled automations ([pricing](https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus)), and IronClaw's hosted Starter plan is shown as free with $5 of credits ([IronClaw](https://www.ironclaw.com/)). ## Voyager is a desktop AI agent for creative work Describe what you want to make; Voyager plans it, uses your apps and tools, and keeps everything editable. [Sign up](/sign-up)