The latest articles on DEV Community by TOPHET (@tophet_cafd932aa262f4fc92). https://dev.to/tophet_cafd932aa262f4fc92 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4033503%2Fdeff0c4f-043f-461d-8858-7deedae15a33.jpg https://dev.to/tophet_cafd932aa262f4fc92 en TOPHET Fri, 17 Jul 2026 09:16:13 +0000 https://dev.to/tophet_cafd932aa262f4fc92/agentic-browser-98-fewer-tokens-than-html-for-llm-web-agents-python-mcp-1onj https://dev.to/tophet_cafd932aa262f4fc92/agentic-browser-98-fewer-tokens-than-html-for-llm-web-agents-python-mcp-1onj <p>Agentic Browser is an <strong>agent-first Python browser</strong> built on Playwright/Chromium so LLMs can drive the web with <strong>compact observations</strong>, <strong>stable element refs</strong>, and <strong>outcome-verified actions</strong> ? not raw HTML dumps.</p> <h2> Why it exists </h2> <p>Traditional scrapers hand models 100k+ tokens of markup. Agents need:</p> <ol> <li>Small structured observations (roles, labels, refs)</li> <li>Actions that mean success (URL/DOM outcomes)</li> <li>A plug-in for any host (MCP + OpenAI/Anthropic tool schemas)</li> </ol> <h2> Measured token efficiency </h2> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Scenario</th> <th>Raw HTML</th> <th>Compact observation</th> <th>Reduction</th> </tr> </thead> <tbody> <tr> <td>Quotes scrape</td> <td>~2.8k?6.2k</td> <td>~0.45k?1.3k</td> <td><strong>~78?84%</strong></td> </tr> <tr> <td>Rockstar GTA VI landing</td> <td>~225,000</td> <td>~1,300</td> <td><strong>~99.4%</strong></td> </tr> <tr> <td>GitHub <code>vercel/next.js</code> </td> <td>~110,000</td> <td>~1,900</td> <td><strong>~98.3%</strong></td> </tr> </tbody> </table></div> <h2> Features </h2> <ul> <li>Stable refs + scoped grounding</li> <li>Outcome verification (e.g. Issues click only OK if URL is <code>/issues</code>)</li> <li>Page gates for challenges (detect &amp; report ? not a bypass tool)</li> <li> <strong>MCP server</strong> for Cursor / Claude Desktop</li> <li> <code>tools_as_openai()</code> / <code>tools_as_anthropic()</code> </li> <li>118 automated tests; milestones M1?M10</li> </ul> <h2> Install </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>pip <span class="nb">install </span>agent-browser playwright <span class="nb">install </span>chromium agent-browser <span class="nt">--help</span> <span class="c"># MCP</span> python <span class="nt">-m</span> agent_browser.mcp </code></pre> </div> <h2> Links </h2> <ul> <li>GitHub: <a href="https://github.com/applejuice093/Agentic-browser" rel="noopener noreferrer">https://github.com/applejuice093/Agentic-browser</a> </li> <li>Release: <a href="https://github.com/applejuice093/Agentic-browser/releases/tag/v0.4.0" rel="noopener noreferrer">https://github.com/applejuice093/Agentic-browser/releases/tag/v0.4.0</a> </li> </ul> <p>MIT ? Python 3.11+</p> python ai llm mcp