<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/">
<channel>
<title>Hronaut blog</title>
<link>https://hronaut.dev/blog</link>
<atom:link href="https://hronaut.dev/blog/rss.xml" rel="self" type="application/rss+xml"/>
<description>Notes on building and using Hronaut, a visible desktop browser for coding agents.</description>
<language>en</language>
<lastBuildDate>Wed, 18 Nov 2026 00:00:00 GMT</lastBuildDate>
<item>
<title>Browser session persistence for Cursor's agent, free with the built-in browser, and when Hronaut is worth paying for</title>
<link>https://hronaut.dev/blog/browser-session-persistence-for-cursor</link>
<guid isPermaLink="true">https://hronaut.dev/blog/browser-session-persistence-for-cursor</guid>
<pubDate>Wed, 18 Nov 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>cursor</category>
<category>browser</category>
<category>mcp</category>
<description>Cursor's built-in browser already keeps the agent signed in between sessions, per workspace. Here is how that works, where it stops, and the cases where a separate browser app like Hronaut earns its $4 a month.</description>
<content:encoded><![CDATA[<p>You want Cursor&#39;s agent to test a page behind a login without signing in every time. Good news first: Cursor already does this for free. This post shows what it keeps, where it stops, and when a separate browser is worth it. I build Hronaut, the paid option at the end, so the free route comes first.</p>
<h2 id="the-free-way-cursors-built-in-browser">The free way: Cursor&#39;s built-in browser</h2>
<p>Cursor ships a browser the agent can drive with no setup. It opens in a separate window or as a pane inside Cursor, and the agent sees its console and network log. Ask for it in chat:</p>
<pre><code class="hljs">@browser Open http://localhost:3000/login, wait for me to sign in,
then open Settings and report any console errors.
</code></pre><p>Cursor&#39;s documentation says browser state persists between agent sessions, per workspace:</p>
<ul>
<li><strong>Cookies</strong>, so authentication and session data stay available.</li>
<li><strong>localStorage and sessionStorage.</strong></li>
<li><strong>IndexedDB.</strong></li>
</ul>
<p>Each workspace has its own browser context, so two projects do not share cookies or storage. Sign in once in a project, and the next agent session in that project is still signed in.</p>
<p>If you work in Cursor, test one account per project, and only Cursor&#39;s agent needs the browser, this is enough. Persistence does not require Hronaut.</p>
<h2 id="a-second-free-option-playwright-mcp">A second free option: Playwright MCP</h2>
<p>If you prefer an external browser, add Microsoft&#39;s open-source <a href="https://github.com/microsoft/playwright-mcp">Playwright MCP</a> in <strong>Cursor Settings › MCP › Add new MCP Server</strong>, as a command server running <code>npx @playwright/mcp@latest</code>. It also keeps a persistent profile by default, and <code>--user-data-dir</code> chooses where.</p>
<h2 id="where-the-free-routes-stop">Where the free routes stop</h2>
<ul>
<li><strong>One browser context per workspace.</strong> The built-in browser keeps one set of cookies per project. Checking the same app as an admin and as a regular member at the same time, or keeping two client accounts apart inside one project, needs a second context you set up yourself.</li>
<li><strong>The browser belongs to Cursor.</strong> Its sign-ins are not available to Claude Code, Codex or another agent working on the same project, and its tabs live with the Cursor window.</li>
<li><strong>No pause.</strong> You can watch the agent, but nothing stops its next action while you type a password or a 2FA code.</li>
</ul>
<h2 id="when-hronaut-is-worth-it">When Hronaut is worth it</h2>
<p>Hronaut is a desktop browser that runs as its own app. Cursor connects to it as an MCP server, and so can Claude Code, Codex and other clients.</p>
<ul>
<li><strong>Several named workspaces per project.</strong> Each has its own cookies and sign-ins, so <code>staging-admin</code> and <code>staging-member</code> can be open side by side and the agent can compare what each one sees.</li>
<li><strong>One sign-in, several agents.</strong> A workspace you signed in to from Cursor can be resumed from Claude Code or Codex with its ID and private resume key.</li>
<li><strong>The browser outlives the editor.</strong> Close Cursor and the workspace stays open with its tabs and page state.</li>
<li><strong>Pause and take over.</strong> Pause agent access, sign in or pass 2FA yourself in the same tab, and hand it back. The agent can ask for your attention on a tab when it reaches a human-only step.</li>
</ul>
<p>It costs $4 a month or $24 a year per named user after a 10-day trial, which needs no card and never charges you. It is one more desktop app to keep updated, and it drives Chromium only.</p>
<h2 id="try-it">Try it</h2>
<p>Install Hronaut from the <a href="https://hronaut.dev/download">download page</a> and add it to Cursor with the install link or configuration on the <a href="https://hronaut.dev/cursor-browser-mcp">Cursor setup page</a>. Then ask:</p>
<pre><code class="hljs">Create Hronaut workspaces named staging-admin and staging-member.
Open https://staging.example.com in each and ask for my attention
so I can sign in as an admin and as a member. Then open the
Billing page in both and tell me what each one shows.
</code></pre><p>The <a href="https://hronaut.dev/blog/persistent-browser-for-claude-code">Claude Code version of this post</a> covers Playwright MCP&#39;s profile in more detail, and the <a href="https://hronaut.dev/blog/persistent-browser-for-codex">Codex version</a> covers Codex&#39;s built-in browser.</p>
<p><strong>Sources:</strong> <a href="https://cursor.com/docs/agent/tools/browser">Cursor Browser documentation</a> (session persistence, per-workspace isolation, pane or window), <a href="https://github.com/microsoft/playwright-mcp">Playwright MCP README</a> (Cursor setup, persistent profile, <code>--user-data-dir</code>), <a href="https://github.com/hronaut/hronaut">Hronaut README</a> (workspaces, pause and takeover, resume), <a href="https://hronaut.dev/pricing">hronaut.dev/pricing</a>, <a href="https://hronaut.dev/cursor-browser-mcp">Cursor setup guide</a>.</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>A persistent browser for Claude Code, free with Playwright MCP, and when Hronaut is worth paying for</title>
<link>https://hronaut.dev/blog/persistent-browser-for-claude-code</link>
<guid isPermaLink="true">https://hronaut.dev/blog/persistent-browser-for-claude-code</guid>
<pubDate>Wed, 18 Nov 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>claude-code</category>
<category>playwright</category>
<category>mcp</category>
<description>You can keep Claude Code signed in to your app between sessions for free with Playwright MCP. Here is the setup, what it does not keep, and the cases where a separate browser app like Hronaut earns its $4 a month.</description>
<content:encoded><![CDATA[<p>You ask Claude Code to check a page behind a login. It opens a browser, hits the sign-in form, and asks you for help. Tomorrow it does the same thing again. This post shows how to stop that with a free tool, and then where the free tool stops. I build Hronaut, the paid option at the end, so the comparison tries hard to give the free path its due.</p>
<h2 id="the-free-way-playwright-mcp">The free way: Playwright MCP</h2>
<p><a href="https://github.com/microsoft/playwright-mcp">Playwright MCP</a> is Microsoft&#39;s open-source MCP server for driving a browser. Add it to Claude Code with one command:</p>
<pre><code class="hljs language-bash">claude mcp add playwright npx @playwright/mcp@latest
</code></pre><p>Start a new Claude Code session and ask it to open your app. A visible browser window opens. Sign in there yourself, once.</p>
<p>That is all it takes, because Playwright MCP keeps a persistent profile by default. Cookies, local storage and sign-ins are saved to disk and come back in the next session. The profile is chosen per project: its folder name includes a hash of the MCP client&#39;s workspace root, so two repositories get two separate profiles and their logins do not mix.</p>
<ul>
<li><strong>Linux:</strong> <code>~/.cache/ms-playwright/mcp-{channel}-{workspace-hash}</code></li>
<li><strong>macOS:</strong> <code>~/Library/Caches/ms-playwright/mcp-{channel}-{workspace-hash}</code></li>
<li><strong>Windows:</strong> <code>%USERPROFILE%\AppData\Local\ms-playwright\mcp-{channel}-{workspace-hash}</code></li>
</ul>
<p>Two flags are worth knowing:</p>
<ul>
<li><code>--user-data-dir=&lt;path&gt;</code> puts the profile where you want it, for example one profile per client account.</li>
<li><code>--isolated</code> keeps the profile in memory only, which is what you want for tests that must start clean. <code>--storage-state=&lt;file&gt;</code> loads saved cookies into such a session.</li>
</ul>
<p>If one project, one account and one Claude Code session at a time is your situation, you can stop reading here. Persistence does not require Hronaut.</p>
<h2 id="what-the-free-way-does-not-keep">What the free way does not keep</h2>
<p>The profile survives. The browser does not. Playwright MCP runs as part of your Claude Code session, and the browser it launched goes away with it:</p>
<ul>
<li><strong>Open tabs and page state.</strong> The next session navigates back to the page and loses anything held only in memory or session storage.</li>
<li><strong>The evidence.</strong> The console and network log, and the screen that showed the bug, are gone once the session ends. Only the agent&#39;s report remains.</li>
</ul>
<p>There are two more limits:</p>
<ul>
<li><strong>One browser per profile.</strong> Playwright&#39;s README says a persistent profile can be used by only one browser instance at a time. Two Claude Code sessions in the same project, or Claude Code and Cursor on the same profile, conflict. The workaround is a separate <code>--user-data-dir</code> or <code>--isolated</code> for each extra client, which also means a separate sign-in.</li>
<li><strong>No pause.</strong> The window is visible and you can type in it, but nothing stops the agent&#39;s next tool call while you enter a password or a 2FA code. You coordinate in chat.</li>
</ul>
<p>More on why the first limit matters is in <a href="https://hronaut.dev/blog/agent-browser-should-outlive-the-session">Why your agent&#39;s browser should outlive the session</a>.</p>
<h2 id="when-hronaut-is-worth-it">When Hronaut is worth it</h2>
<p>Hronaut is a desktop browser that runs as its own app. Claude Code connects to it over a local MCP endpoint, and so can Codex, Cursor and other clients. That changes what survives:</p>
<ul>
<li><strong>The browser outlives the session.</strong> Close Claude Code and the workspace stays open with its tabs, sign-ins and page state. A new session returns to the same workspace with its ID and private resume key.</li>
<li><strong>Named workspaces per project or client.</strong> Each workspace has its own cookies and site storage, managed in one window instead of profile folders. Several agents can work in separate workspaces at once, and one agent at a time holds write access to each.</li>
<li><strong>Pause and take over.</strong> You can pause agent access, sign in or pass 2FA yourself in the same tab, and hand it back. The agent can ask for your attention on a tab when it reaches a step only a person should do.</li>
<li><strong>One workspace, several clients.</strong> A workspace Claude Code signed in to can be resumed from Codex or Cursor without signing in again.</li>
</ul>
<p>It costs $4 a month or $24 a year per named user after a 10-day trial, which needs no card and never charges you. It is also one more desktop app to install and keep updated, and it drives Chromium only.</p>
<h2 id="try-both-on-the-same-task">Try both on the same task</h2>
<p>Start with Playwright MCP. If you notice you keep losing tabs, run two sessions against the same app, or want to stop the agent while you sign in, install Hronaut and connect Claude Code:</p>
<pre><code class="hljs">/plugin marketplace add hronaut/hronaut-claude-plugin
/plugin install hronaut@hronaut
</code></pre><p>Then give the same task:</p>
<pre><code class="hljs">Create a Hronaut workspace named staging-admin with scratch storage.
Open https://staging.example.com and stop at the sign-in form.
Ask for my attention on that tab, wait for me to sign in,
then open Settings and report what you see.
</code></pre><p>Close Claude Code, start a new session, and ask it to resume <code>staging-admin</code>. The tab is where you left it. The <a href="https://hronaut.dev/setup">setup page</a> has configurations for other clients, and the <a href="https://hronaut.dev/browser-mcp-guide">browser MCP guide</a> compares Hronaut with Playwright MCP, Chrome DevTools MCP and extension-based tools in more detail.</p>
<p><strong>Sources:</strong> <a href="https://github.com/microsoft/playwright-mcp#user-profile">Playwright MCP README</a> (install command, persistent profile, locations, one instance per profile, <code>--isolated</code>, <code>--storage-state</code>), <a href="https://github.com/hronaut/hronaut">Hronaut README</a> (workspaces, pause and takeover, write lease, resume), <a href="https://hronaut.dev/pricing">hronaut.dev/pricing</a>, <a href="https://hronaut.dev/setup">hronaut.dev/setup</a> (Claude Code plugin).</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>A persistent browser for Codex, free with the built-in browser or Playwright MCP, and when Hronaut is worth paying for</title>
<link>https://hronaut.dev/blog/persistent-browser-for-codex</link>
<guid isPermaLink="true">https://hronaut.dev/blog/persistent-browser-for-codex</guid>
<pubDate>Wed, 18 Nov 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>codex</category>
<category>playwright</category>
<category>mcp</category>
<description>Codex can keep you signed in to your app between tasks for free. Here are the three free routes, the built-in browser, the Chrome extension and Playwright MCP, what each keeps, and the cases where Hronaut earns its $4 a month.</description>
<content:encoded><![CDATA[<p>You ask Codex to check a page behind a login, and it stops at the sign-in form. Codex has more than one way to avoid that, and most of them are free. This post goes through them in order. I build Hronaut, the paid option at the end, so the free routes come first and get a fair description.</p>
<h2 id="free-route-1-the-built-in-browser-in-the-chatgpt-desktop-app">Free route 1: the built-in browser in the ChatGPT desktop app</h2>
<p>If you run Codex in the ChatGPT desktop app, it already has a browser. Open a tab with <strong>New tab</strong> or Ctrl+T (Cmd+T on macOS) and point the agent at it with <code>@Browser</code>.</p>
<p>OpenAI&#39;s documentation describes it as a browser profile separate from your regular browser. It does not share your existing tabs or sign-ins, you can sign in directly when a task needs an account, and it keeps its own profile and browsing history. <strong>Settings › Browser</strong> is where you clear that data.</p>
<p>For one app and one account this is the simplest answer: sign in once in the built-in browser and let Codex work there.</p>
<h2 id="free-route-2-the-chrome-extension">Free route 2: the Chrome extension</h2>
<p>The ChatGPT browser extension lets Codex work in an existing Chrome, Edge, Brave, Opera or Vivaldi tab with your regular browser profile, invoked with <code>@Chrome</code>. Everything you are already signed in to is available at once.</p>
<p>That is also its cost. The agent acts inside the browser that holds your email, your bank and your other accounts. Use it for tasks where that is what you want, and keep it away from tasks that do not need your identity.</p>
<h2 id="free-route-3-playwright-mcp-in-codex-cli">Free route 3: Playwright MCP in Codex CLI</h2>
<p>In the terminal, Codex CLI has no built-in browser, but it speaks MCP. Add Microsoft&#39;s open-source <a href="https://github.com/microsoft/playwright-mcp">Playwright MCP</a> server:</p>
<pre><code class="hljs language-bash">codex mcp add playwright npx <span class="hljs-string">&quot;@playwright/mcp@latest&quot;</span>
</code></pre><p>Playwright MCP keeps a persistent profile by default, so cookies and sign-ins survive between Codex sessions. The profile folder is named after a hash of the MCP client&#39;s workspace root. To be sure each repository gets its own sign-ins, pass a profile per project:</p>
<pre><code class="hljs language-bash">codex mcp add playwright-myapp -- npx <span class="hljs-string">&quot;@playwright/mcp@latest&quot;</span> --user-data-dir=<span class="hljs-string">&quot;<span class="hljs-variable">$HOME</span>/.codex/browser-profiles/myapp&quot;</span>
</code></pre><p><code>--isolated</code> keeps the profile in memory for tests that must start clean.</p>
<h2 id="where-the-free-routes-stop">Where the free routes stop</h2>
<ul>
<li><strong>One identity per profile.</strong> The built-in browser has one profile, and a Playwright profile holds one set of sign-ins. Checking a page as an admin and as a regular member at the same time, or keeping two clients&#39; accounts apart in one project, needs two profiles you set up and keep track of yourself.</li>
<li><strong>Your everyday profile, or nothing.</strong> The extension shares your real sign-ins. The other two start empty. Neither gives you a middle ground of named, separate profiles you can see and manage in one place.</li>
<li><strong>The browser belongs to one tool.</strong> The built-in browser lives in the ChatGPT app, and a Playwright MCP browser closes with the Codex CLI session that launched it, along with its tabs and console. A sign-in made there is not available to Claude Code or Cursor.</li>
<li><strong>No pause.</strong> You can watch, but nothing stops the agent&#39;s next action while you type a password or a 2FA code.</li>
</ul>
<h2 id="when-hronaut-is-worth-it">When Hronaut is worth it</h2>
<p>Hronaut is a desktop browser that runs as its own app, next to Codex. Codex CLI and the ChatGPT desktop app connect to it as an MCP server.</p>
<ul>
<li><strong>Named workspaces.</strong> One per project, client or role, each with its own cookies and sign-ins, managed in one window. <code>staging-admin</code> and <code>staging-member</code> can be open side by side.</li>
<li><strong>Your everyday browser stays out.</strong> Hronaut never touches your Chrome profile.</li>
<li><strong>The browser outlives the session.</strong> Close Codex and the workspace stays open with its tabs and page state. The next session resumes it with its ID and private resume key, and Claude Code or Cursor can resume the same workspace without signing in again.</li>
<li><strong>Pause and take over.</strong> Pause agent access, sign in or pass 2FA yourself in the same tab, and hand it back. The agent can ask for your attention on a tab when it reaches a human-only step.</li>
</ul>
<p>It costs $4 a month or $24 a year per named user after a 10-day trial, which needs no card and never charges you. It is one more desktop app to keep updated, and it drives Chromium only.</p>
<h2 id="try-it">Try it</h2>
<p>Install Hronaut from the <a href="https://hronaut.dev/download">download page</a>, then add the plugin in a terminal:</p>
<pre><code class="hljs language-bash">codex plugin marketplace add hronaut/hronaut-codex-plugin
codex plugin add hronaut@hronaut
</code></pre><p>Start a new Codex session and give it a task that needs a sign-in:</p>
<pre><code class="hljs">Create a Hronaut workspace named staging-admin with scratch storage.
Open https://staging.example.com and stop at the sign-in form.
Ask for my attention on that tab, wait for me to sign in,
then open Settings and report what you see.
</code></pre><p>The <a href="https://hronaut.dev/codex-browser-mcp">Codex setup guide</a> covers manual configuration and checks, and the <a href="https://hronaut.dev/blog/persistent-browser-for-claude-code">Claude Code version of this post</a> has more on what Playwright MCP keeps.</p>
<p><strong>Sources:</strong> <a href="https://developers.openai.com/codex/app/browser">ChatGPT desktop browser documentation</a> (separate profile, sign-in, Settings › Browser, <code>@Browser</code> and <code>@Chrome</code>), <a href="https://github.com/microsoft/playwright-mcp#user-profile">Playwright MCP README</a> (Codex command, persistent profile, <code>--user-data-dir</code>, <code>--isolated</code>), <a href="https://github.com/hronaut/hronaut">Hronaut README</a> (workspaces, pause and takeover, resume), <a href="https://hronaut.dev/pricing">hronaut.dev/pricing</a>, <a href="https://hronaut.dev/codex-browser-mcp">Codex setup guide</a> (plugin commands).</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>The image that never loaded</title>
<link>https://hronaut.dev/blog/page-that-never-finishes-loading</link>
<guid isPermaLink="true">https://hronaut.dev/blog/page-that-never-finishes-loading</guid>
<pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>electron</category>
<category>debugging</category>
<category>mcp</category>
<description>One image whose server never answered held every script-based call an agent made in its tab. The cause is a wait inside Electron; the fix is to wait for the document, not for the whole page.</description>
<content:encoded><![CDATA[<p>On October 22, 2026, an agent in Hronaut sat on “Running a script on the page” for minutes. The page looked finished. One image on it was not: its server had accepted the request and never answered.</p>
<h2 id="the-symptom">The symptom</h2>
<p>Every agent call that runs a script in the page waited until the agent&#39;s request timed out: reading the page (<code>browser_snapshot</code>), clicking (<code>browser_click</code>), running code (<code>browser_evaluate</code>). The workspace stayed busy for the agent&#39;s next calls, and the tab kept showing the agent at work.</p>
<h2 id="the-cause">The cause</h2>
<p>Hronaut runs its page scripts through Electron&#39;s <code>webContents.executeJavaScript</code>. While the main frame is loading, Electron does not run the script. It waits:</p>
<pre><code class="hljs language-ts"><span class="hljs-keyword">const</span> <span class="hljs-title function_">waitTillCanExecuteJavaScript</span> = <span class="hljs-keyword">async</span> (<span class="hljs-params">webContents</span>) =&gt; {
  <span class="hljs-keyword">if</span> (webContents.<span class="hljs-title function_">getURL</span>() &amp;&amp; !webContents.<span class="hljs-title function_">isLoadingMainFrame</span>()) <span class="hljs-keyword">return</span>;
  <span class="hljs-keyword">return</span> <span class="hljs-keyword">new</span> <span class="hljs-title class_">Promise</span>(<span class="hljs-function">(<span class="hljs-params">resolve</span>) =&gt;</span> {
    webContents.<span class="hljs-title function_">once</span>(<span class="hljs-string">&#x27;did-stop-loading&#x27;</span>, <span class="hljs-function">() =&gt;</span> { <span class="hljs-title function_">resolve</span>(); });
  });
};
</code></pre><p><code>did-stop-loading</code> fires when the tab&#39;s loading spinner stops. An image that never answers keeps the spinner going, so the script never starts. The DevTools console doesn&#39;t wait like this: it runs your code as soon as the document is there.</p>
<h2 id="the-fix">The fix</h2>
<p>Hronaut now decides when a page script may run. A script runs once the shown document is ready (its <code>DOMContentLoaded</code>), as in DevTools, while images and other resources keep loading. Navigating and waking a sleeping tab go on after 30 seconds while the page keeps loading, as in Chrome. A page that has shown nothing after 30 seconds gets a plain error, “Timed out waiting for the page to finish loading”, instead of a request that hangs.</p>
<p>One more trap turned up on the way. A navigation that ends with a 204 response or a download fires no event that says it ended, so Hronaut gives the new document three seconds to appear after the request ends, then treats the old page as current again.</p>
<p>The fix shipped in Hronaut 3.5.0, on October 22, 2026.</p>
<h2 id="what-to-take-from-it">What to take from it</h2>
<p>If you build agent tools on Electron or another embedded browser, “the page finished loading” is the wrong gate for page scripts. Long polling, a stalled CDN, an analytics beacon: real pages often never finish. Gate on the document, cap every wait, and tell the agent what happened instead of letting its request time out.</p>
<p><strong>Sources:</strong> <a href="https://github.com/electron/electron/blob/main/lib/browser/api/web-contents.ts">Electron, <code>lib/browser/api/web-contents.ts</code></a> (<code>waitTillCanExecuteJavaScript</code>), <a href="https://www.electronjs.org/docs/latest/api/web-contents#event-did-stop-loading">Electron docs: <code>did-stop-loading</code></a>, <a href="https://github.com/hronaut/hronaut/blob/main/CHANGELOG.md">Hronaut changelog, 3.5.0</a>, <a href="https://github.com/hronaut/hronaut/blob/main/src/main/browser/page-script-readiness.ts">Hronaut source: <code>page-script-readiness.ts</code></a>, <a href="https://github.com/hronaut/hronaut/pull/1186">pull request #1186</a>.</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>Why your agent's browser should outlive the session</title>
<link>https://hronaut.dev/blog/agent-browser-should-outlive-the-session</link>
<guid isPermaLink="true">https://hronaut.dev/blog/agent-browser-should-outlive-the-session</guid>
<pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>mcp</category>
<category>architecture</category>
<category>browser</category>
<description>An agent session lasts minutes. The browser it works in holds sign-ins, tabs and evidence that took longer to build. Three lifetimes to keep apart, and when a short-lived browser is the right call.</description>
<content:encoded><![CDATA[<p>Many browser tools for coding agents start a browser when the agent connects and close it when the agent is done. For automated tests that is the right default. For everyday work next to a person, it throws away the part that took longest to build.</p>
<h2 id="what-closes-with-the-session">What closes with the session</h2>
<p>Some tools keep a profile on disk, so sign-ins can survive: Playwright MCP does this by default. What still goes away with the browser process:</p>
<ul>
<li><strong>Open tabs and live page state.</strong> The next session reopens and re-navigates the pages under test, and anything kept only in session storage or in memory is gone.</li>
<li><strong>The evidence.</strong> The console, the network log and the screen that showed the problem.</li>
<li><strong>A view for the person.</strong> If no one was watching, the only record is the agent&#39;s own report.</li>
</ul>
<p>The last one matters more than it sounds. An agent can report “done” while the page shows a login wall or an empty shell. A browser that is still open, and visible, after the session ends is a cheap way to check.</p>
<h2 id="three-clocks">Three clocks</h2>
<table>
<thead>
<tr>
<th>Lifetime</th>
<th>Typical length</th>
<th>What it holds</th>
</tr>
</thead>
<tbody><tr>
<td>Agent task</td>
<td>minutes</td>
<td>one prompt, thread or job</td>
</tr>
<tr>
<td>MCP connection</td>
<td>a request, or a client process</td>
<td>messages between the agent and a tool server</td>
</tr>
<tr>
<td>Browser</td>
<td>days or weeks, if you let it</td>
<td>process, profile, tabs, sign-ins, windows</td>
</tr>
</tbody></table>
<p>With the stdio transport, the client starts the MCP server as a subprocess and ends it, so a browser owned by that server ends with the client. With Streamable HTTP, the server runs on its own and each request gets its own response, so the browser can outlive any one client. The transport doesn&#39;t decide ownership, but it makes one choice easier than the other.</p>
<h2 id="keeping-data-is-not-keeping-the-browser">Keeping data is not keeping the browser</h2>
<p>A persistent profile answers “will the cookies still be there?” It doesn&#39;t answer who owns the browser, or whether two clients can use it at once. Playwright MCP&#39;s README says it plainly: a persistent profile can only be used by one browser instance at a time, so parallel clients in the same project need <code>--isolated</code> or a separate profile directory.</p>
<h2 id="when-a-short-lived-browser-is-right">When a short-lived browser is right</h2>
<p>CI, regression suites, and anything that must start clean, run in parallel on servers and leave nothing behind. There, a browser that dies with the task is a feature, and a test runner such as Playwright is the right tool. Hronaut is not built for that.</p>
<h2 id="how-hronaut-keeps-the-browser">How Hronaut keeps the browser</h2>
<ul>
<li>The desktop app owns Chromium and stays open while agents connect and disconnect.</li>
<li>Each task can work in its own named workspace with persistent storage, so parallel tasks don&#39;t share sign-ins.</li>
<li>An agent that comes back resumes its workspace with a private key. The workspace ID alone grants no access.</li>
<li>A person can pause new commands, take over a tab and hand it back.</li>
</ul>
<p>The longer version, with failure modes and four patterns compared, is <a href="https://hronaut.dev/coding-agent-browser-lifecycle">Why browser lifecycle matters for coding agents</a>.</p>
<p><strong>Sources:</strong> <a href="https://github.com/microsoft/playwright-mcp#user-profile">Playwright MCP README, “User profile”</a>, <a href="https://modelcontextprotocol.io/specification/2026-07-28/basic/transports">MCP specification (2026-07-28): transports</a>, the <a href="https://hronaut.dev/coding-agent-browser-lifecycle">browser lifecycle guide</a>, and <a href="https://hronaut.dev/blog/why-i-built-my-own-browser-for-ai-coding-agents">Why I built my own browser for AI coding agents</a> (an agent reporting “done” on a login page).</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>Let your coding agent check your localhost app</title>
<link>https://hronaut.dev/blog/check-localhost-with-a-coding-agent</link>
<guid isPermaLink="true">https://hronaut.dev/blog/check-localhost-with-a-coding-agent</guid>
<pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>localhost</category>
<category>testing</category>
<category>mcp</category>
<description>Install Hronaut, connect your agent, and have it test your dev server in a visible browser that stays put while you edit and restart. Five steps and one prompt.</description>
<content:encoded><![CDATA[<p>Your coding agent can change the code. Checking the result is the slow part: you open the page, click through the flow, read the console, and tell the agent what broke. Here is how to let the agent do that check itself, in a browser window you can watch.</p>
<p>You need a local coding agent that speaks MCP, such as Claude Code, Codex or Cursor, and a dev server you can start.</p>
<h2 id="1-install-and-start-hronaut">1. Install and start Hronaut</h2>
<p>Download the package for your system from the <a href="https://hronaut.dev/download">download page</a> and start it. Hronaut is a desktop app for Windows, macOS 13 or newer, and Linux. The macOS builds are signed and notarized by Apple, and the download page shows how to check the file against its published hash and GitHub build attestation. The first 10 days are free, counted from your agent&#39;s first tool call. The trial needs no card and never charges you.</p>
<h2 id="2-connect-your-agent">2. Connect your agent</h2>
<p>In Claude Code:</p>
<pre><code class="hljs">/plugin marketplace add hronaut/hronaut-claude-plugin
/plugin install hronaut@hronaut
</code></pre><p>In a terminal, for Codex CLI or the ChatGPT desktop app:</p>
<pre><code class="hljs language-bash">codex plugin marketplace add hronaut/hronaut-codex-plugin
codex plugin add hronaut@hronaut
</code></pre><p>The plugins read Hronaut&#39;s MCP token on their own, so there is nothing to paste. Other agents have a ready configuration on the <a href="https://hronaut.dev/setup">setup page</a> and on Hronaut Home. Start a new session and run <code>/mcp</code> in Claude Code or Codex: Hronaut should show as connected.</p>
<h2 id="3-start-your-dev-server-on-a-fixed-address">3. Start your dev server on a fixed address</h2>
<p>For Vite: <code>npx vite --host 127.0.0.1 --port 5173 --strictPort</code>. Browsers keep local storage and IndexedDB per origin, that is per scheme, host and port, so <code>http://localhost:5173</code>, <code>http://127.0.0.1:5173</code> and <code>http://localhost:5174</code> do not share them. <code>--strictPort</code> makes Vite stop with an error instead of quietly moving to the next free port.</p>
<h2 id="4-ask-for-the-check">4. Ask for the check</h2>
<p>Paste this, with your own address and finish line:</p>
<pre><code class="hljs">Create a Hronaut workspace named local-check with scratch storage.
Open http://127.0.0.1:5173 and wait for the page heading.
Submit the sign-up form empty, then with a valid email.
Report what the page shows each time, console errors,
and failed network requests.
</code></pre><p>The agent creates an isolated workspace and opens the page in a visible tab. It reads the page as an accessibility snapshot, clicks and types by element, and reads the console and network log. You watch the same tab and can pause the agent or take over at any time.</p>
<h2 id="5-edit-restart-check-again">5. Edit, restart, check again</h2>
<p>Fix what the agent found, restart the server if needed, and ask it to check again in the same tab. Cookies and storage live in the workspace, not in your dev server, so they survive the restart. If the server is down for a moment, Hronaut keeps the failed address in the tab and the agent can retry it.</p>
<p>A new agent session the next day can return to the same workspace with its ID and private resume key, instead of starting over.</p>
<h2 id="where-this-stops">Where this stops</h2>
<ul>
<li>Hronaut drives Chromium on your computer. It is not a CI runner and does not cover Firefox or WebKit. Turn each bug it finds into a test in your repository.</li>
<li>The agent must run on the same machine. A cloud agent sees its own localhost, not yours: see <a href="https://hronaut.dev/client-topology">connection troubleshooting</a>.</li>
<li>Keep the dev server on 127.0.0.1. There is no need to expose it with <code>0.0.0.0</code> or a tunnel for this.</li>
</ul>
<p>More detail is in the <a href="https://hronaut.dev/coding-agent-localhost-browser">localhost guide</a>.</p>
<p><strong>Sources:</strong> <a href="https://hronaut.dev/setup">hronaut.dev/setup</a> (plugin commands), the <a href="https://hronaut.dev/coding-agent-localhost-browser">localhost guide</a> (origins, restarts, limits), <a href="https://vite.dev/config/server-options">Vite server options</a> (<code>strictPort</code>), <a href="https://developer.mozilla.org/en-US/docs/Glossary/Origin">MDN: origin</a>.</p>
<p><em>Written with the help of AI.</em></p>
]]></content:encoded>
</item>
<item>
<title>Why I built my own browser for AI coding agents in 2026</title>
<link>https://hronaut.dev/blog/why-i-built-my-own-browser-for-ai-coding-agents</link>
<guid isPermaLink="true">https://hronaut.dev/blog/why-i-built-my-own-browser-for-ai-coding-agents</guid>
<pubDate>Sat, 24 Oct 2026 00:00:00 GMT</pubDate>
<dc:creator>Yevhen Tienkaiev</dc:creator>
<category>ai</category>
<category>mcp</category>
<category>testing</category>
<category>webdev</category>
<description>Our team wanted an AI agent to test our platform like a real user. No existing browser tool fit, so I built Hronaut. What was missing, how it works, and its limits.</description>
<content:encoded><![CDATA[<p><em>Originally published on <a href="https://dev.to/hronom/why-i-built-my-own-browser-for-ai-coding-agents-in-2026-m3i">DEV Community</a>.</em></p>
<p>At my day job we needed manual testing for our platform: acting like a real user under different conditions, such as a slow network, different screen sizes, and signed-in flows. It is 2026, so the obvious idea was to let an AI agent walk through the site and check it instead of a person.</p>
<p>We looked at the existing tools. There are many, and none of them fit what we needed. So, oddly enough for 2026, I ended up writing my own app: a separate, visible browser that an AI agent drives. That became <a href="https://hronaut.dev/">Hronaut</a>.</p>
<p>This post covers what the existing tools were missing, how the browser is built, and what I learned along the way.</p>
<h2 id="what-existing-tools-were-missing">What existing tools were missing</h2>
<p>We wanted a browser where the agent checks the site like a real user, while a human stays in control. That gave us a list of requirements:</p>
<ul>
<li><strong>Visible.</strong> A person sees what the agent is doing and can step in.</li>
<li><strong>The browser outlives the agent.</strong> If the agent session drops or restarts, the tabs and sign-ins stay.</li>
<li><strong>Isolated.</strong> Each task gets its own profile, and nothing touches the developer&#39;s personal browser.</li>
<li><strong>A human signs in.</strong> The agent should never know passwords, but it should be able to continue in a signed-in session.</li>
<li><strong>Local.</strong> Test localhost and internal staging without moving anything to the cloud.</li>
<li><strong>Any agent.</strong> Claude Code, Codex, Cursor, anything that speaks MCP.</li>
</ul>
<p>Each existing option covered some of these, never all at once. Playwright MCP is a great automation tool, but its browser is designed for a robot, not for working next to a human. Chrome extensions hand the agent your real browser with all your personal accounts. Cloud browsers are great for scale, but they can&#39;t simply reach a local dev server.</p>
<p>On our platform, this got very concrete. With Playwright MCP, the browser closed together with the agent session, so after every restart we signed in again and reopened the pages under test. When the agent ran in a browser nobody was watching, the only way to know what it had done was to read the logs afterwards. Sometimes it reported &quot;done&quot; while it was actually stuck on the login page or looking at an empty page. Trying a Chrome extension meant handing the agent my real work browser, with my email and personal accounts, and I wasn&#39;t willing to do that. And a cloud browser couldn&#39;t open localhost or our internal staging environment at all.</p>
<h2 id="design-principles">Design principles</h2>
<ol>
<li><strong>The browser is always visible.</strong> The agent works in a normal Chromium window. You watch it live instead of reading a report afterwards.</li>
<li><strong>The human outranks the agent.</strong> One click pauses it; you can take over and hand control back. When the agent needs a person (to sign in, pass a check, or confirm an action), Hronaut sends a notification and marks the tab.</li>
<li><strong>The browser doesn&#39;t depend on the agent session.</strong> The agent disconnects, the browser keeps running. The agent comes back and picks up the same tabs.</li>
<li><strong>One task, one workspace.</strong> Its own cookies, storage, and sign-ins. Several agents can work in parallel without getting in each other&#39;s way.</li>
<li><strong>Passwords are not for the agent.</strong> A human signs in, and everything the agent reads from the page goes through a filter that hides passwords, API keys, session identifiers, and signed URLs.</li>
<li><strong>Bring your own agent and model.</strong> Hronaut is only the browser, and it runs on your machine, not in the cloud.</li>
</ol>
<h2 id="how-it-works">How it works</h2>
<p>Hronaut is an Electron desktop app for Windows, macOS, and Linux. Inside it runs an MCP server (Streamable HTTP transport) that listens only on <code>127.0.0.1</code> (port 47812 by default). The agent connects to it like any other MCP server and gets tools to open a tab, navigate, click, type, read the page, and take a screenshot.</p>
<p>A few decisions I&#39;m happy with:</p>
<ul>
<li><strong>Workspaces with a stable ID.</strong> The agent creates a workspace and gets a private resume key. After a disconnect or restart it returns to its own workspace and can&#39;t see anyone else&#39;s.</li>
<li><strong>A copy instead of access to the original.</strong> A person signs in once, and the agent works in a copy with those cookies and localStorage. The original stays untouched, even when the agent has no direct access to it.</li>
<li><strong>Exclusive write access.</strong> Only one client writes to a workspace at a time; anyone can read. Two agents won&#39;t click the same button at once.</li>
<li><strong>Observer workspaces.</strong> A clean, signed-out profile tied to one site, where the agent can only look. Handy for checking what a page looks like to a stranger.</li>
<li><strong>Page quality checks.</strong> The agent can ask whether the page has real content or is an empty shell, a login wall, a soft 404, or someone else&#39;s domain. Fewer false &quot;done&quot; reports.</li>
<li><strong>Token auth on by default.</strong> The MCP endpoint requires a token, stored in a file only the owner can read. Setup instructions in the app read the token from that file, so it never lands in an agent&#39;s config or in the chat.</li>
</ul>
<h2 id="try-it">Try it</h2>
<ol>
<li>Download Hronaut for your OS from the <a href="https://github.com/hronaut/hronaut/releases/latest">releases page</a> and start it.</li>
<li>Connect your agent:<ul>
<li>Claude Code: run <code>/plugin marketplace add hronaut/hronaut-claude-plugin</code>, then <code>/plugin install hronaut@hronaut</code>. The plugin reads the token from Hronaut&#39;s token file on each connect.</li>
<li>Codex and other clients: copy the ready-made setup from the Home page inside Hronaut; it already includes the token setting.</li>
</ul>
</li>
<li>Check the connection with <code>/mcp</code> in your agent, then ask: &quot;Create a Hronaut workspace called first-check, open <a href="https://example.com">https://example.com</a> in it, and tell me the page title.&quot;</li>
</ol>
<p><em>Updated October 28, 2026: the Claude Code plugin moved to its own repository, so the marketplace command now names <code>hronaut/hronaut-claude-plugin</code>. Current setup for every agent is on the <a href="https://hronaut.dev/setup">setup page</a>.</em></p>
<h2 id="limitations">Limitations</h2>
<ul>
<li><strong>Paid after the trial.</strong> The first 10 days are free with no card, counted from the agent&#39;s first call; after that it needs a subscription. If free Playwright MCP covers your automated tests without sign-ins, you probably don&#39;t need Hronaut.</li>
<li><strong>Not open source.</strong> The code is readable on <a href="https://github.com/hronaut/hronaut">GitHub</a>, but under its own license.</li>
<li><strong>Local only.</strong> The agent has to run on the same machine. For hundreds of browsers on servers you need a cloud service.</li>
<li><strong>Bring your own agent and model.</strong> Hronaut doesn&#39;t write tests by itself.</li>
<li><strong>Verify the download.</strong> Hashes and GitHub build attestations are published for every package. (Since November 2026 the macOS builds are signed and notarized by Apple.)</li>
</ul>
<h2 id="wrapping-up">Wrapping up</h2>
<p>My main takeaway: an AI agent in a browser needs more than access to pages. It needs a person nearby who can see what&#39;s happening and step in. That&#39;s what Hronaut is for.</p>
<p>How do you give your agent a browser? Do you trust it with your own Chrome, keep a separate profile, or not let it in at all? I&#39;d love to hear your stories and criticism in the comments, and I&#39;ll answer questions about how Hronaut works myself.</p>
<p><em>Disclosure: I&#39;m the developer of Hronaut. This article was written with the help of AI.</em></p>
]]></content:encoded>
</item>
</channel>
</rss>
