Local-first web agents: natural-language goal in, real browser actions, schema-validated JSON out.https://garrison-hq.github.io/enhttps://garrison-hq.github.io/sortie/blog/introducing-sortie/https://garrison-hq.github.io/sortie/blog/introducing-sortie/sortie is an open-source, local-first platform for autonomous web agents — natural-language goal in, real browser actions, schema-validated JSON out. AGPL-licensed, runs entirely on your machine. Thu, 11 Jun 2026 00:00:00 GMT<p>A <strong>sortie</strong> is a mission launched out from a garrison: go out, do the job, come back. That’s exactly what this tool does — you give it a goal, it ventures out onto the web in a real browser, and it comes back with structured data.</p> <p>sortie is an open-source, local-first platform for autonomous web agents. Describe the data you want in plain language plus a schema, and sortie drives a real Chromium browser, locates content by <em>meaning</em> rather than brittle CSS/XPath selectors, and returns clean, schema-validated JSON. For multi-step tasks — “log in, search for X, collect the first 20 results” — an agent loop plans and executes navigation, clicks, typing, and pagination, recovering from failures along the way.</p> <div><h2 id="why-local-first">Why local-first</h2></div> <p>Plenty of hosted services will run a browser fleet for you. sortie makes the opposite bet: <strong>everything runs on your machine.</strong> Your sessions, your credentials, your scraped data — none of it transits someone else’s cloud.</p> <p>That choice shapes the security model:</p> <ul> <li><strong>Credentials never reach the model.</strong> Goals and traces only ever contain <code dir="auto">{{cred:NAME}}</code> placeholders; substitution happens at the moment of typing, observations are scrubbed, and values are never persisted.</li> <li><strong>Login sessions stay on disk.</strong> Named profiles are stored at <code dir="auto">0600</code> permissions, never enter the database, and are never returned by the API.</li> <li><strong>Bring your own LLM.</strong> Anthropic, OpenAI, or any OpenAI-compatible endpoint — including a fully local Ollama or vLLM. With a local model and a SearXNG sidecar, the whole loop runs without a single external API call.</li> </ul> <p>And one boundary worth stating plainly: <strong>CAPTCHA solving and anti-bot evasion are out of scope, permanently.</strong> Runs that hit such walls fail gracefully with a clear reason. Be polite to the sites you automate.</p> <div><h2 id="whats-in-the-box">What’s in the box</h2></div> <ul> <li><strong>CLI</strong> — <code dir="auto">extract</code>, <code dir="auto">agent</code>, <code dir="auto">search</code>, <code dir="auto">fetch</code>, saved <code dir="auto">query</code> replays, login <code dir="auto">profile</code> management, concurrent <code dir="auto">batch</code> runs, and <code dir="auto">runs</code> history/export.</li> <li><strong>Typed SDK</strong> — zod schemas in, inferred types out.</li> <li><strong>REST + WebSocket server</strong> with a playground UI: author a run, watch it execute step by step with screenshots.</li> <li><strong>MCP server</strong> — give Claude Code (or any MCP-capable agent) the web as an API: <code dir="auto">web_search</code>, <code dir="auto">web_fetch</code>, <code dir="auto">web_extract</code>, <code dir="auto">run_agent</code>, and more.</li> </ul> <div><h2 id="open-source-agpl">Open source, AGPL</h2></div> <p>sortie is <a href="https://github.com/garrison-hq/sortie/blob/main/LICENSE">AGPL-3.0</a>: use it, modify it, self-host it freely — and if you run a modified version as a service, share your changes under the same terms.</p> <p>The code is on GitHub at <a href="https://github.com/garrison-hq/sortie">garrison-hq/sortie</a>. Start with the <a href="https://garrison-hq.github.io/sortie/getting-started/">getting started guide</a>, and come say hi in the issues if you build something with it — or if something breaks.</p>announcement