# Dejo > Dejo is a free, open-source (MIT) browser automation MCP server that remembers websites. It gives AI agents such as Claude Code, Codex and Cursor a real browser: the agent learns a website task once, and every run after that is one tool call with zero model calls. When a site changes, Dejo repairs only the step that moved. For AI agents: the index of this site is https://dejo.run/llms.txt, and every page is also served as markdown (add .md to its address, or send Accept: text/markdown). Key facts: made by Rohit Kashyap and first released on 15 October 2026. Free, with no paid plan, no account and no API key: Dejo never calls a model, the user's own agent does the thinking. It runs on the user's machine (Python 3.11 or newer; Windows, macOS, Linux) and keeps its memory in a local SQLite file, ~/.dejo/memory.db, never passwords. Five MCP tools: dejo_run does every website task; dejo_login, dejo_profile, dejo_memory and dejo_manage handle sign-in, accounts and memory. For developers who use AI coding agents on websites. Contact: https://github.com/dejohq/dejo-mcp/issues Dejo remembers how to do tasks on websites, not just your login. That is the difference from Playwright MCP and Chrome DevTools MCP, which start from nothing in every new session. Use it for website tasks an agent will do again; for a one-off look at a single page it saves nothing. When a site changes between runs: every saved step is checked after it runs and the replay stops at the exact step that did not land. A moved button is found with another of up to nine saved ways; a toast, newsletter box or cookie banner in the way is closed safely (reject before accept, never buy, pay, subscribe or sign up); a different signed-in account stops the run before step 1; only the broken step is learned again. When the task changes (0.5.1): new values are swapped in or the run stops, "this month" follows today's date, another subject or deed is never replayed as the same job, and each profile keeps its own routes. Details: /docs/when-the-site-changes ## Benchmark On 19 tasks on real, lesser-known websites, each in a brand-new Claude session (Claude Sonnet 5, medium effort), Dejo's first run used 38% fewer tokens than Playwright MCP (251K against 409K) and answered 19 of 19 right (Playwright MCP 17, Playwright CLI 18). Every run after replayed from memory with 78% fewer tokens (89K), in a single tool call on 17 of 19 tasks. On a separate 50-site reliability test with the same fixed steps for each tool and no model, Dejo 0.4.0 found and used the right element on 47 of 50 sites, plain Playwright on 45. These are our own measurements; the method and raw data are public. Dejo is not cheaper on single-page, one-line lookups. ## Core pages - [Dejo: browser automation MCP that remembers websites](https://dejo.run/index.md): Dejo is an open-source browser automation MCP server for Claude Code, Codex and Cursor. It learns a website once, then replays it with zero model calls. - [Playwright MCP token usage: measured on 19 real tasks](https://dejo.run/benchmark.md): 19 tasks on lesser-known websites, a fresh Claude session each: Dejo used 38% fewer tokens than Playwright MCP on the first run and 78% fewer on every run after, with every answer right. The method, the raw data and the honest limits. - [Use cases for a browser MCP that remembers websites](https://dejo.run/use-cases.md): Weekly numbers, end-to-end tests that check every step, iPhone checks after each deploy, several signed-in accounts, admin work and price checks: website tasks your AI agent repeats, learned once and replayed with zero model calls. - [Free tools for AI agents and browser automation](https://dejo.run/tools.md): Free tools for developers who use AI coding agents and automate browsers: MCP config generator, Claude Code cost calculator, Playwright locator generator, selector tester, token cost calculator and more. Nothing you type leaves your browser. - [MCP config generator: Claude Code, Codex, Cursor](https://dejo.run/tools/mcp-config-generator.md): Describe your MCP server once and copy the exact config for Claude Code (claude mcp add or .mcp.json), Codex (config.toml), Cursor, VS Code and Claude Desktop. Free, runs in your browser, checked against each client's docs. - [Claude Code cost calculator (2026 prices)](https://dejo.run/tools/claude-code-cost-calculator.md): How much does Claude Code cost? Pro $20, Max $100 or Ultra $200 a month, or about $13 per developer per active day on API billing (Anthropic's figure). Compare plans and API for your team, free. - [Playwright locator generator: HTML to getByRole](https://dejo.run/tools/playwright-locator-generator.md): Paste an element's HTML and get its Playwright locators, best first: getByRole with the accessible name, getByLabel, getByPlaceholder, getByText, getByTestId, in JavaScript or Python. Plus an XPath to CSS converter. Free, in your browser. - [XPath tester and CSS selector tester online](https://dejo.run/tools/xpath-css-selector-tester.md): Paste HTML and test an XPath or CSS selector: see every element it matches, with a count. Uses your browser's own engines, so results match a real page. Free, nothing is sent anywhere. - [AI token cost calculator: Claude vs OpenAI prices](https://dejo.run/tools/ai-token-cost-calculator.md): Enter tokens in and out and calls per month, and compare the cost on every current Claude and OpenAI model, with prompt caching. Paste text to estimate its tokens. Official prices, free, in your browser. - [AI crawler checker: GPTBot, ClaudeBot and robots.txt](https://dejo.run/tools/ai-crawler-checker.md): Type a domain and see which AI crawlers its robots.txt allows (GPTBot, OAI-SearchBot, ClaudeBot, Google-Extended, PerplexityBot and more) and whether it has an llms.txt file. Free robots.txt checker for AI. - [llms.txt generator: make one for your website](https://dejo.run/tools/llms-txt-generator.md): Fill in your site's name, summary and main pages and copy a ready llms.txt in the llmstxt.org format: H1, summary, link sections and an Optional section. Free, in your browser. - [User agent parser: what is my user agent?](https://dejo.run/tools/user-agent-parser.md): Read any user agent string: browser and version, engine, operating system, device, and whether it is a crawler such as GPTBot, ClaudeBot or Googlebot. Yours is filled in. Free, in your browser. ## Docs - [Dejo MCP server setup: install in a minute](https://dejo.run/docs.md): Install the Dejo MCP server with pip, add it to Claude Code, Codex or Cursor, and run your first website task. Free and open source; it takes about a minute. - [Browser automation in Claude Code: add a browser MCP](https://dejo.run/docs/claude-code.md): Give Claude Code a browser that remembers websites: one command to add the Dejo MCP server, then what to ask. - [Browser automation in Codex: the config.toml setup](https://dejo.run/docs/codex.md): Give OpenAI Codex a browser MCP server that remembers websites: the config.toml block and how to check it works. - [Cursor browser automation with a browser MCP](https://dejo.run/docs/cursor.md): Add the Dejo browser MCP server to Cursor with one mcp.json entry, then let Cursor's agent use websites. - [Keep your AI agent signed in to websites](https://dejo.run/docs/sign-in.md): Sign in once yourself, in a real window. Dejo keeps the session, so your agent starts signed in every time and never sees your password. - [When a website changes: how Dejo replays safely](https://dejo.run/docs/when-the-site-changes.md): Every saved step is checked. Dejo finds a moved button another way, clears pop-ups and cookie banners safely, stops on the wrong account, and relearns only the step that broke. - [Dejo MCP tools reference](https://dejo.run/docs/tools.md): Every tool the Dejo MCP server gives your agent, what it does, and when the agent should call it. - [Troubleshooting Dejo](https://dejo.run/docs/troubleshooting.md): Fixes for the common problems: the agent does not see Dejo, the browser is missing, a profile is stuck, or a site changed. ## Compare - [Playwright MCP alternative that remembers websites](https://dejo.run/compare/playwright-mcp.md): Dejo is a Playwright MCP alternative for website tasks you repeat: on 19 tasks across 11 real websites it used 38% fewer tokens on the first run and 78% fewer on every run after. Plus Playwright CLI, Chrome DevTools MCP and Claude in Chrome compared. - [Chrome DevTools MCP vs Playwright MCP (measured)](https://dejo.run/compare/chrome-devtools-mcp.md): Chrome DevTools MCP debugs a page, Playwright MCP drives one. Both start from nothing every session. We measured both against Dejo over 90 fresh Claude sessions. - [Browser Use alternative: Dejo vs browser-use](https://dejo.run/compare/browser-use.md): browser-use is a browser agent with its own model; Dejo is a browser MCP with no model that remembers each task for Claude Code, Codex or Cursor. What each does, side by side, and when to pick which. - [Stagehand alternative: Dejo vs Stagehand](https://dejo.run/compare/stagehand.md): Stagehand is Browserbase's SDK where a model finds elements and caches them on its servers; Dejo is a local browser MCP with no model that replays whole tasks for your agent. Side by side, and when to pick which. - [Browser MCP alternative: Dejo vs Browser MCP](https://dejo.run/compare/browser-mcp.md): Browser MCP puts your agent in your own open Chrome through an extension; Dejo gives it its own browser that remembers each task and replays it in one call. Side by side, from browsermcp.io's own words. - [Dejo vs agent-browser (Vercel), compared](https://dejo.run/compare/agent-browser.md): agent-browser is Vercel's Rust CLI that drives a browser one command at a time; Dejo is an MCP server that remembers each website task and replays it in one call. Side by side, and when to pick which. ## Blog - [Blog](https://dejo.run/blog.md): Measured notes on browser automation for AI agents: what Claude Code, Codex and Cursor pay to use websites, and how to make them remember. - [Codex browser: what works in the CLI and IDE](https://dejo.run/blog/codex-browser.md): OpenAI's built-in Browser runs in ChatGPT, not in the Codex CLI or IDE extension. How to give Codex a browser there with an MCP server in config.toml, and which one to pick. - [Playwright MCP login: keep your agent signed in](https://dejo.run/blog/playwright-mcp-login.md): Playwright MCP keeps you logged in with a persistent profile per project. Where it lives, how to sign in without giving the AI your password, and saving the login to a file with --storage-state. - [Browser agent token statistics (2026, measured)](https://dejo.run/blog/browser-agent-token-statistics.md): Citable numbers from 19 real-website tasks in Claude Code: tokens, calls, cost, seconds and right answers for Playwright MCP, Playwright CLI and Dejo, what each turn costs, and site memory. Every number links its raw data. - [Why Playwright MCP uses so many tokens](https://dejo.run/blog/why-playwright-mcp-uses-so-many-tokens.md): Measured in 19 Claude Code sessions: every turn re-sends about 39,700 tokens before any page is read, and snapshots are smaller than you think. The bill follows the number of turns, from 164K to 1.36M per task. - [Site memory: a new task on a known site, 33% cheaper](https://dejo.run/blog/browser-agent-site-memory.md): We gave Claude Code a new task on 10 real websites, with and without Dejo remembering the site from a different task. With the site known it averaged 238K tokens instead of 354K, every answer right. Where it didn't help, too. - [Best browser MCP for Claude Code (2026, measured)](https://dejo.run/blog/best-browser-mcp-claude-code.md): Chrome DevTools MCP, Playwright MCP, Playwright CLI, Claude in Chrome and Dejo: which to pick for debugging, tests, your own Chrome or repeated tasks, with measured Claude Code numbers. - [Playwright MCP vs Claude in Chrome: which to use](https://dejo.run/blog/playwright-mcp-vs-claude-in-chrome.md): Claude in Chrome works in your signed-in Chrome while you watch; Playwright MCP runs its own browser for tests and CI. Measured Playwright MCP numbers, Anthropic's docs, and what neither remembers. - [Playwright MCP vs Playwright CLI vs memory (measured)](https://dejo.run/blog/playwright-mcp-vs-cli.md): 19 real-site tasks in fresh Claude Code sessions: Playwright MCP averaged 409K tokens, Playwright CLI 559K, and a browser with memory 251K once, then 89K. Which to pick. - [Test your website as an iPhone from Claude Code](https://dejo.run/blog/test-website-iphone-claude-code.md): Have Claude Code open your site as an iPhone 13, screenshot it and check the layout, then repeat the check after every deploy with zero model calls. Real runs on five pages. - [Run a website task every day with zero AI cost](https://dejo.run/blog/schedule-browser-task-zero-ai-cost.md): Let your AI agent learn a browser task once, then let cron or Windows Task Scheduler run it every day with no model calls. Real output, exit codes and scripts. - [Why browser agents fail on real websites: 7 causes](https://dejo.run/blog/why-browser-agents-fail.md): Late content, slow confirmations, iframes, cookie banners, dialogs, ambiguous selectors and captchas: what breaks AI browser agents, and what we changed in Dejo for each. - [AI agent with multiple signed-in user roles](https://dejo.run/blog/ai-agent-multiple-user-roles.md): How to let an AI agent act as several users on one site without signing out and back in, compared with Playwright's saved login files. - [Can Claude log in to websites? Three ways that work](https://dejo.run/blog/can-claude-log-in-to-websites.md): Claude can read public links, but signing in takes a browser. Claude in Chrome, Playwright MCP and Dejo each keep you signed in differently: what each does with your login, and which to pick. - [Self-healing Playwright: repair one step, not the run](https://dejo.run/blog/self-healing-browser-automation.md): Two kinds of self-healing: Playwright's healer agent fixes the test file, fallback locators fix the step at run time. How each works, and why every step needs a check. - [Browser agent memory: why agents re-read websites](https://dejo.run/blog/browser-agent-memory.md): Every new chat, your AI agent learns the same website again. We measured what that costs over 90 Claude sessions, and what changes when the browser remembers the route. ## Machine-readable - [Full content](https://dejo.run/llms-full.txt): every page above in one file - [Markdown catalog](https://dejo.run/api/md/_catalog): every markdown route and how to fetch it - [Sitemap](https://dejo.run/sitemap.xml): every HTML page - [AI usage policy](https://dejo.run/ai.txt): what AI systems may and may not do with this content - [Agent skill](https://dejo.run/.well-known/agent-skills/dejo/SKILL.md): how an agent installs and uses Dejo - [Docs MCP server](https://dejo.run/mcp): a remote, read-only MCP server (Streamable HTTP, no key) with tools to list, search and read these pages; details at https://dejo.run/mcp/server-card. It serves the docs; Dejo itself runs locally - [OpenAPI](https://dejo.run/openapi.json): the public endpoints of this site; all of them in https://dejo.run/.well-known/api-catalog ## Optional - [Changelog](https://dejo.run/changelog.md): Every release of Dejo, the browser automation MCP server that remembers websites: what changed and how to upgrade. - [Privacy](https://dejo.run/privacy.md): What the Dejo software and the dejo.run website collect. The software sends nothing to us; the website uses analytics only with cookies you allow. - [Terms](https://dejo.run/terms.md): The terms for the Dejo software (MIT licensed, open source) and the dejo.run website. - [GitHub](https://github.com/dejohq/dejo-mcp): source code, MIT license - [PyPI](https://pypi.org/project/dejo-mcp/): the dejo-mcp package