An open-source desktop console that runs your own local AI agent inside managed browser profiles, so multi-account social work stays isolated, scheduled and observable.
Used this tool? Rate it
Used this tool? Rate it
Nextbrowser is a desktop console that supervises browser work done by an AI agent. It is not a chatbot with a browser bolted on, and — despite what most third-party listings still say — it is not a cloud service. The application is an Electron, React and TypeScript desktop program for macOS and Windows, published as open source under the GNU Affero General Public License v3.0 only.
The distinction that matters most is where the intelligence comes from. Nextbrowser ships no model of its own. The official documentation is explicit that the selected agent still needs to be installed and authenticated on your computer, and it names two supported runtimes: Claude Code, which connects through its CLI installation, and Codex, which connects through the executable bundled with the ChatGPT desktop app. You bring the agent and its subscription; Nextbrowser brings the browser context, the identity management and the observation surface around it.
That division of labour explains the product's whole shape. A prompt alone does not make browser automation work. Someone has to choose which browser identity to use, keep that identity separated from a dozen others, start and stop the session, notice when a page stalls, and stop a run that has drifted from its instructions. Nextbrowser's own framing is that browser work by an AI agent spans more than a prompt, and that an operator must choose a browser identity, control the session, keep the agent process observable, and recover when a page or run fails. The desktop console is where those controls live.
If you research this product anywhere other than its own site, you will probably be told it is dead. That impression is dated, and understanding why matters before you evaluate anything else.
Nextbrowser began as a cloud-based social media management platform with credit-based subscriptions. On 27 April 2026, founder Stan Sadokov announced the shutdown of that platform, citing unsustainable infrastructure costs, a codebase that needed a full rewrite, and a market too small to justify solving those problems. Coverage of that announcement is still the top search result for the product's name, and competing vendors have written comparison articles around it.
The company then rebuilt. In its own words, the team paused the cloud platform, spent months testing anti-detect browsers, agent runtimes and open-weight models, and concluded that the missing piece was not another browser or another agent but an orchestration layer above the tools that already existed. The rebuilt product is the desktop application described here. The public repository was created in July 2026 and has shipped roughly thirty releases since; version 0.2.4 was published on 10 August 2026, with commits landing the same day.
So the accurate status is: the original cloud product was discontinued, and a differently-architected open-source desktop product now carries the name. Both facts are true, and reading only one of them will mislead you. Directory entries that describe Nextbrowser as a closed-source cloud service with monthly credits are describing a product that no longer exists, and the pricing page those entries link to now returns a 404.
The desktop console organises everything around a small number of layers, each of which the operator controls directly.
Profiles as browser identities. A profile represents a browser identity and its associated state, letting you separate browser contexts for different tasks, accounts, customers, regions or test runs. The documentation is careful about what a profile does not guarantee: selecting one does not by itself prove that its session is running, that its expected tab is active, or that its identity matches the requested country. Those must be verified.
Sessions. A session is the running browser context for a profile, and it must be running before an agent can work with its pages. Session diagnostics cover status, open tabs, page state, screenshots and identity verification.
Proxy and fingerprint rotation. Rotation requests a refreshed browser identity, optionally pinned to a country, and the docs recommend verifying the result afterwards whenever a workflow depends on geography. The framing here is unusually honest for this category: rotation is described as a recovery and isolation tool, not a promise that a site will accept a session, because sites apply their own policies and risk systems.
The agent workspace. Around whichever local runtime you connect, the console provides chat history and named conversations, queued prompts and run status, stop and edit controls, conversation forks, and streamed output. One caveat is documented plainly: Nextbrowser passes the selected profile name to the local agent as context, but this is guidance, not technical containment — you are told to confirm the actual profile and running session yourself.
Skills and custom scripts. A skill is a reusable instruction set for a domain or browser workflow; a custom script is a private equivalent for work too specific to publish. Before dispatch, a preflight sequence can confirm the profile is running, start it if needed, inspect open tabs, activate or open the target page, wait for readiness, and then send the prepared prompt. Preflight establishes a starting context; it does not guarantee the page stays unchanged or that the workflow succeeds.
Scheduled runs. A schedule stores a prompt, a selected agent, a time, a weekday selection and an enabled state. The critical constraint, repeated across the documentation and the use-case pages, is that scheduled runs execute locally while the desktop app is open. Nextbrowser is not a cloud service that keeps working after you close your laptop.
Live View and diagnostics. Live View exposes the running browser page inside the console, complementing structured diagnostics rather than replacing them.
Captcha tooling. The app can detect a challenge and invoke an available handling path. The documentation refuses to oversell this: a captcha is controlled by the site or its provider, so detection or a solve attempt can fail, be throttled, require a human, or be disallowed for the task, and the tools should never be described as a universal bypass.
The marketing surface targets one audience unambiguously: people who operate many social accounts at once. The published use-case playbooks cover Reddit account incubation, social media monitoring, account warm-up, influencer outreach, SEO link-building, and LLM optimisation — testing prompts and tracking how brands appear across AI assistants. The warm-up playbook lists Instagram, YouTube, X, Reddit, TikTok, LinkedIn, Facebook and Snapchat as supported platforms.
Growth marketers and agencies running dozens of profiles are the primary fit. The company's own account of its rebuild describes the target scenario as a marketer growing fifty Instagram accounts or running eighty Reddit accounts, each needing isolated profiles, clean fingerprints per session and geo-matched proxies.
A second and quieter audience is anyone who wants supervised agentic browsing at all. If you already pay for Claude Code or a ChatGPT desktop subscription and want that agent to work inside a browser you can watch, pause and inspect, the console delivers that irrespective of whether you manage a single account or fifty.
The product is least suited to developers who want a programmable cloud API. The current release is a human-operated desktop console. The old cloud platform exposed APIs and webhooks for programmatic session management; that surface went away with it, and the rebuilt application has not advertised a replacement.
Setup is deliberately short, and each step exists because skipping it causes a failure that looks like something else later.
.dmg and .zip) and Windows x64 (.exe).The habit worth forming early is diagnosing from the outside in. The documented order is account connection, proxy availability, profile, session, active tab, page state, identity verification, agent installation, then workflow instructions. Stopping at the first failed check prevents a stopped session or a missing tab from being misread as an agent failure ten minutes later.
Turning a task into a repeatable workflow. Review the skill or custom script and strip out secrets and stale assumptions. Choose the profile, target site and agent deliberately. Let preflight establish the browser context. Then supervise the run, stopping or editing when the task changes, and verify the final page before accepting any consequential action.
Rotating identity mid-workflow. Stop or preserve in-progress work that depends on the current identity, request rotation with an explicit country if geography matters, verify the resulting identity, then reopen or re-check the target page. Rotation can invalidate login state or change what a site presents, so the documentation warns against rotating blindly during a transaction.
Keeping a human in the loop. This is stated as an operating rule rather than a suggestion: purchases, publishing, account changes, deletions and other consequential actions should stay under your control. Two related warnings are worth internalising — prompt limits are instructions rather than enforced approval gates, and page content should be treated as untrusted input, because a page can contain instructions intended to redirect an agent. That last point is a genuine prompt-injection risk inherent to any agent that reads live web pages.
The application itself is open-source software under AGPL-3.0-only, and the builds are downloadable from GitHub without a purchase. The licence permits commercial use, modification and redistribution, with the copyleft condition that if you distribute a modified version — or make one available for remote network interaction — you must offer the corresponding source under the same licence.
Beyond that, the honest answer about cost is that there is no public pricing page at the time of writing. The path that third-party articles still link to returns a 404, and the current site's navigation offers Blog, Use Cases, FAQ, Docs and Download with no pricing entry. The historical credit-based tiers described in competitor comparisons belonged to the discontinued cloud platform and should not be treated as current.
What the documentation does make clear is that free software does not mean zero total cost. Three cost centres sit outside the download:
It runs only while your computer does. Scheduled runs execute locally while the app is open. Anyone migrating from the old cloud product, or from a competitor's hosted service, is giving up unattended 24/7 execution.
Platform coverage is stated inconsistently. The July 2026 rebuild announcement says the product is available for Mac, Windows and Linux. The repository README, the product guide and the actual release assets all describe macOS and Windows only, with no Linux build published. Treat the documentation and the release page as authoritative, and do not plan a Linux deployment on the strength of the blog post. On macOS, published binaries currently target Apple Silicon; Intel Mac users should verify availability before committing.
The version number is a warning label. At v0.2.4 with roughly thirty releases in about six weeks, this is early software moving fast toward a stated v1.0.0 goal. Recent release notes describe fixes for white screens and chat-mode replay bugs. Expect rough edges and frequent updates.
Adoption signals are thin. The public repository shows a small number of stars and forks, and download counts on the newest assets are in single digits. Independent validation is correspondingly limited: G2 and Product Hunt pages could not be retrieved for this review, and targeted community searches surfaced no substantive third-party discussion of the rebuilt application. The testimonials on the site are the vendor's own — and one of them is attributed to a director of Eyes of Wonder, the company that operates the product, without that relationship being disclosed on the page.
The advertised use cases carry real account risk. Warming accounts, operating dozens of profiles per platform and posting from them can violate the terms of service of the platforms involved, whatever tooling is used. Nextbrowser's own terms place that responsibility on you: you are responsible for complying with applicable laws, platform rules, account agreements and data protection obligations, and the product must not be used for unauthorised access, credential theft, spam, abuse, malware or activity intended to bypass restrictions unlawfully. Fingerprint rotation and residential proxies reduce correlation signals; they do not make a policy violation permissible, and the documentation never claims a site will accept a session.
Two document dates do not line up. The Terms of Use carry an effective date of 16 July 2026, consistent with the rebuild. The Privacy Policy is dated 29 July 2025, a year earlier, and still describes the cloud-era architecture. Where the two conflict on how data is handled, note that the desktop application's behaviour is documented separately in the repository, and the older policy has not been visibly refreshed for the new architecture.
For a browser product this is the section that decides adoption, and it is worth separating three distinct data paths.
Who operates it. The Privacy Policy discloses that the product is operated by Eyes of Wonder Ltd and developed by NodeMaven FZ-LLC. Naming both an operator and a developer entity is more disclosure than many tools in this category offer. NodeMaven is itself known as a proxy provider, which is consistent with proxy access being bundled into the account layer.
What the desktop app reports. The repository documents the telemetry precisely. The desktop build enables a configured GA4 stream by default, and events include an application instance identifier and operational metadata. Critically, the documentation states that the analytics implementation does not send prompt text, target URLs, or page domains. The published event list is consistent with that claim: it covers app lifecycle, agent connection and turn outcomes, session and profile actions, skill and schedule usage — and where prompt size appears at all, it appears as a length bucket rather than content. Collection can be disabled for a build by setting a blank measurement ID. Because the client is open source, these claims are auditable rather than merely asserted.
What the platform stores. The Privacy Policy states that the service collects account registration data including an email address and a hashed password, along with user input such as task names, configurations, workflows and uploaded scripts, plus automatically collected device, usage and inferred-location data via Amplitude, Google Analytics and Google Tag Manager. Data is stored on AWS in the us-east-1 region, with encryption in transit and at rest and encrypted backups. The policy states that personal data is not sold or rented, that data may be shared with infrastructure, analytics, support and legal recipients, and that anonymised or aggregated data may be used for research and product improvement.
On training. The policy lists its purposes as providing and improving the service, authenticating accounts, executing and storing automated browser tasks, analysing usage, preventing misuse and meeting legal obligations, and it says anonymised or aggregated data may be used for research, analytics and product improvement. It does not state that customer content is used to train AI models — but it also contains no explicit commitment that it will not be. Since the models are supplied by whichever agent you connect, the more consequential training question is governed by Anthropic's or OpenAI's terms for your own subscription, not by Nextbrowser. Anyone with strict data-handling requirements should confirm both sides in writing.
Your rights. The policy enumerates access, rectification, erasure, restriction, portability, objection and consent withdrawal, exercised through a published legal contact address. It also disclaims responsibility for data handled by external scripts or by web pages visited during task execution — a reasonable boundary, but one that means the sites your agent visits are outside this policy's protection.
Against hosted browser-automation APIs such as Browserbase or Apify, Nextbrowser trades programmability and unattended execution for local control, source availability and the absence of a per-session cloud bill. If you need a scriptable endpoint that runs without a human present, a hosted platform remains the right category.
Against conventional anti-detect browsers such as Multilogin, GoLogin, Dolphin Anty or AdsPower, the pitch is that fingerprint isolation alone was never the hard part. The company's argument is that those tools handled fingerprinting but still required doing everything manually inside them, and that the missing layer was orchestration. Nextbrowser positions itself above that layer rather than as a replacement for it, and states that existing browser, proxy and agent setups are intended to plug into the orchestration layer — though that interoperability is described as work in progress.
Against general-purpose agentic browsers, the differentiator is multi-account discipline: per-account fingerprinting, per-session proxy rotation, warming schedules and platform-specific detection patterns are exactly the specifics the team says general agents did not handle.
The genuinely distinctive combination is source-available code under a copyleft licence, local execution, bring-your-own-agent architecture, and an explicit telemetry contract. That combination is rare in a category where opacity is the norm.
It is a desktop application for macOS and Windows. The AI agent and the browser session both run on your own computer. An earlier, separate cloud platform under the same name was discontinued in April 2026, which is why many third-party listings still describe a cloud product.
Yes. The application is published under the GNU Affero General Public License v3.0 only, in a public repository, and the licence permits commercial use, modification and redistribution. The copyleft condition is that a modified version you distribute — or expose for remote network interaction — must be offered with corresponding source under the same licence.
None of its own. It connects to an AI agent you have already installed and authenticated: Claude Code through its CLI, or Codex through the executable bundled with the ChatGPT desktop app. Model quality, usage limits and billing come from that agent's own subscription.
The application is free to download under its open-source licence, and there is no public pricing page at present. Costs sit elsewhere: your AI agent subscription, proxy traffic associated with the account layer, and your own hardware. Commercial terms for proxy capacity are not published on the site, so confirm them directly before scaling.
No. Scheduled runs execute locally while the desktop app is open. This is stated in both the documentation and the use-case pages, and it is the clearest functional difference from the discontinued cloud product.
The repository documentation states that the desktop analytics implementation does not send prompt text, target URLs, or page domains; the published event list reflects operational actions, with prompt size appearing only as a length bucket. Analytics can be disabled for a build by setting a blank measurement ID, and because the client is open source the behaviour can be audited. Separately, the platform's Privacy Policy describes collecting account data, task configurations and usage telemetry, stored on AWS in us-east-1.
The tooling provides isolation — separate profiles, per-session proxy and fingerprint rotation, and staged activity — but no tool makes multi-accounting compliant with a platform's terms. The vendor's own terms place responsibility for platform rules and account agreements on you, and the documentation explicitly declines to promise that any site will accept a rotated session.
It can detect a challenge and invoke an available handling path, but the documentation states plainly that a solve attempt can fail, be throttled, require a human, or be disallowed for the task, and that these tools should never be described as a universal bypass.
The repository documentation and the actual release assets list macOS and Windows only. A July 2026 company blog post does mention Linux availability, but no Linux build is published on the releases page, so the documentation and release assets should be treated as authoritative.
Early. The public repository was created in July 2026, the current release is v0.2.4 as of 10 August 2026, and roughly thirty releases have shipped on the way to a stated v1.0.0 target. Development is visibly active, but stars, forks and download counts remain small, and independent third-party validation of the rebuilt application is currently scarce.