OpenClaw Foundation · Alternatives
OpenClaw alternatives, grouped by why you are leaving
Almost everything on this list is free software, because that is what this corner of the market looks like: you pay the model, not the tool. The real choice is not price but shape — a coding agent instead of an assistant, a foundation behind the project instead of nobody, someone else's machine instead of yours, or a workflow that cannot open a shell at all. Four reasons, four different shortlists.
Why people look.
Worth separating, because these four reasons lead to four shortlists that barely overlap — and a generic "best AI assistants" list serves none of them.
You wanted a coding agent
The most common arrival and the easiest to fix. OpenClaw is a personal assistant that can code through a skill, and the popular way it does so is by launching another agent underneath. If the job is writing software, the tools it delegates to are the direct answer, and you lose nothing but a chat-window front end.
The blast radius is too wide
A daemon with shell and filesystem access, reachable from inbound messages, is a large thing to run on a machine that matters. Its own documentation says to treat inbound messages as untrusted and that it is not a security boundary between mutually distrusting users. Narrower tools exist: some ask permission per action, some cannot execute anything at all.
There is nobody to call
No company, no SLA, no support contract, no invoice — a foundation holds the copyright and the licence disclaims warranty. For regulated work or anywhere procurement needs a counterparty, that ends it. The fix is not necessarily a commercial vendor: a project under a recognised foundation is a different governance answer at the same price.
You do not want to run a machine
Self-hosting is the point of OpenClaw, and for some people it is the objection. Keeping a daemon patched across four release channels, deciding on sandboxing and reviewing what the security audit says is unpaid work that arrives every week. Hosted agents move that cost onto somebody else's operations team and charge for it.
If you only read one line.
goose
If you like the shape and want governance behind it. The closest like-for-like on this list: open source, runs on your own machine, works with any provider including subscriptions you already pay for — but developed under the Linux Foundation's agentic-AI foundation rather than by a project with no institution attached.
Claude Code
If the job was always writing code. It is one of the agents OpenClaw's coding skill launches underneath, so going direct removes a layer rather than losing a capability — and it is the tool we have tested most heavily, with published transcripts.
Manus
If you want the assistant but not the server administration. It runs in someone else's cloud, so there is no daemon to patch and no shell of yours to protect. You pay for that in money and in giving up local file access.
Every alternative we checked.
Ordered by how many of the four reasons above each tool answers, not by any score — we run no benchmark that ranks personal assistants against each other, and a number here would be invented. Entry prices are read from each vendor's own page on the date shown in the row; where a tool has its own hub on this site, the row links there instead of repeating a figure that would go stale.
The like-for-like swap: an open-source agent that runs locally and takes any model you have.
Everything structural that makes OpenClaw appealing, with an institution behind it. It is a native desktop app and a CLI, written in Rust, and it explicitly supports bringing subscriptions you already pay for as well as raw API keys — the same money question this site answers on the pricing page. Governance is the real difference: the project moved under the Linux Foundation's agentic-AI foundation, which is a meaningful answer to "who is accountable" that does not cost anything.
It is an agent for work at your desk, not a presence in your messaging apps. If the reason you wanted OpenClaw was answering WhatsApp from your phone while the agent works on your home machine, that is the capability you give up — goose has no equivalent of the channel layer.
Best for people who want a local open-source agent with a foundation standing behind it.
The original "give a model your computer" project, still actively developed.
The closest match to what OpenClaw actually does on your machine: run code, edit documents, drive the browser, operate the filesystem — locally. It predates most of this category, has a desktop app alongside the CLI, and remains a live project rather than a 2023 artefact. If you want the capability without the messaging surface, this is the narrower tool that still does the dangerous part well.
It carries the same fundamental risk as OpenClaw — a model executing code on your computer — so it answers the governance and coding questions but not the blast-radius one. Choosing it because it feels safer would be a mistake; choosing it because it has less attack surface exposed to inbound messages would be correct.
Best for people who want local code execution without wiring an agent into their messaging accounts.
A terminal coding agent from a frontier lab — and one of the two OpenClaw delegates to.
If the task was always software, this removes a layer instead of adding one: OpenClaw's popular coding skill runs Claude Code underneath, so going direct gets the same engine without the daemon, the channels or the plugin. It is also the tool this site has tested most heavily, with capability runs published as transcripts rather than claims, and its costs are priced route by route on our own hub.
It has no free tier whatsoever — the cheapest way in is a paid subscription, which is a real step down from software that costs nothing to install. It is also strictly a coding tool: it will not answer your messages, sort your calendar or watch your inbox.
Best for developers who want the coding engine without the assistant wrapped around it.
The other agent OpenClaw's coding skill launches, and the one with a genuinely free tier.
The same argument as Claude Code with a different lab and a different billing shape, plus one advantage that matters if free software was the appeal: Codex is included on the free ChatGPT plan as limited access rather than a trial, so there is a way in that costs nothing. If you already pay for ChatGPT, OpenClaw can use that plan too — but going direct gets you the cloud-side features that the assistant layer never exposes.
It meters in credits and OpenAI publishes no price for one, so budgeting means deriving what a credit buys — we do exactly that on its pricing page, but a derivation is a harder conversation with finance than a rate. The free tier's limits are also unpublished.
Best for people who want a coding agent with a free way in and already live in ChatGPT.
A provider-agnostic terminal coding agent, MIT-licensed and very widely starred.
The answer for someone who liked that OpenClaw takes any model and wanted that property in a coding tool. It connects to any provider, and like OpenClaw it can sign in with subscriptions you already hold rather than demanding a key. MIT rather than Apache, terminal-first with IDE and desktop surfaces, and one of the most-starred projects in the whole category.
Coding only — no channels, no calendar, no inbox, none of the assistant surface. It also sells a curated model service alongside the free client, so "free" describes the tool while the smoothest path through it is a paid one.
Best for terminal-first developers who want to keep choosing their own models.
A general-purpose agent that runs in the vendor's cloud instead of on your machine.
The straight answer to "I want the assistant, not the sysadmin work". Nothing to install, nothing to patch, no daemon holding your API keys, and no shell of yours for a prompt injection to reach. It runs research, builds documents and executes long multi-step tasks on its own infrastructure, with scheduled and concurrent tasks included on every paid tier.
You give up the two things OpenClaw exists for: local file access and running for free. It is also metered in credits rather than money, which puts it in the same awkward budgeting position as Codex — the monthly allowance is published as a credit count, and what a credit buys is not.
Best for people who want agent work done without operating any infrastructure.
An open-source agent inside the editor, which asks before it acts.
The narrow-blast-radius option for people who still want an agent editing files and running commands. It lives in the editor rather than as a system daemon, it is free for individual developers with no subscription, and it takes your own API keys — so the spending model is the same pay-for-what-you-use shape as OpenClaw on a key, without a background process listening to the internet.
It is bound to the editor, so nothing happens while you are away from your desk — the exact opposite of an assistant answering messages on your behalf. Enterprise pricing is quote-only, so an organisation cannot budget it from the public page.
Best for developers who want an agent they watch, not one that runs unattended.
Workflow automation with optional AI steps — the answer when you never wanted an agent.
A large share of what people ask a personal assistant to do is not judgement at all: move this attachment, file that email, post this summary on a schedule. Those are workflows, and a workflow does exactly what it did yesterday. Self-hosting is free, it connects to the same messaging channels, and an LLM can be one step inside a run rather than the thing holding the keys to your machine.
It cannot improvise, which is the entire point of an agent — anything you did not build in advance simply does not happen, and building it is your work rather than the model's. The self-hosted edition is also fair-code rather than an OSI-approved open-source licence, so "open source" is not quite the right word for it.
Best for recurring, well-defined tasks where predictability beats initiative.
Pick by what you're doing.
It is going on a work laptop
You liked the demo, and the machine you would install it on holds client data or production credentials.
Cline
An editor-bound agent you watch has a fraction of the exposure of a daemon reachable from inbound messages, and it keeps the part you actually wanted — a model that can read the repository and run commands. If it truly has to be OpenClaw, the defensible version is a dedicated machine or virtual server, sandboxing on, and the built-in security audit run after every config change.
Procurement needs a name on the contract
The tool is fine; the answer to "who supports this" is not.
goose
A project under the Linux Foundation's agentic-AI foundation is a governance answer rather than a commercial one, and it often clears review where an unaffiliated project does not — at the same price of nothing. If review demands an actual counterparty with an SLA, no free tool on this page qualifies and the conversation moves to a paid vendor.
You just want your inbox handled
Sorting mail, filing attachments, posting a daily digest — the same tasks every day.
n8n
None of that needs judgement, and a workflow that runs identically every time will not surprise you at three in the morning. Put a model in one step if a summary needs writing. This is the cheapest and by far the least dangerous way to get most of the value people actually describe when they talk about a personal assistant.
You are only there to write software
The assistant framing was incidental — you wanted something to work through a codebase.
Claude Code
Go to the engine directly rather than through a plugin that launches it. Compare it against the alternative on our comparisons, and if a free way in matters more than model choice, Codex is the one with a free tier that is not a trial.
When to stay put.
The honest counterweight — in these cases moving costs more than it saves.
Frequently asked.
What people search next.