Dejo (browser MCP)Partly tested🌏 World

Rohit Kashyap built Dejo, a browser MCP that lets Claude Code learn a website task once and replay it with no model calls; he measured 78% fewer tokens on repeat runs

Driving websites through Playwright MCP means the agent re-reads every page each time, so a task it did yesterday costs full tokens again today. Rohit Kashyap (@rohit_jsfreaky) built Dejo, a browser MCP for agents: the first time, the agent walks the task and Dejo saves each step together with what should be true afterwards; after that the same task is one tool call, and Dejo replays it without reading the page or calling a model. When a site changes, it stops at the step that didn't land and asks the agent to relearn only that step. Against Playwright MCP on 19 tasks across 11 lesser-known sites, he reports 38% fewer tokens on the first run and 78% fewer on every run after, with the data public. We pip-installed dejo-mcp 0.5.1 and dejo doctor passed; acting as the agent through a minimal MCP client, we taught it to read the first paragraph on example.com, then replayed it twice from a fresh process with dejo run: 105 ms and 85 ms, model_calls 0. We did not reproduce his 19-task benchmark.

Visit site Dejo (browser MCP) ↗
Pricing
Free
Region
Global
Where it's from
Overseas
Added
Nov 12, 2026
Tested on
Nov 12, 2026
Version we used
dejo-mcp 0.5.1 from PyPI (repo be4d39a), Playwright 1.65.1, system Chrome
Categories
Automation & AgentsCoding
Original author
Rohit Kashyap @rohit_jsfreaky ↗
Our test
Partly tested
Cost / time
MIT, with no paid plan, account or API key; Dejo never calls a model. The first, learning run is billed by your own agent (Claude, Codex, etc.) as usual.
Source
X @rohit_jsfreaky ↗dejo.run ↗GitHub dejohq/dejo-mcp ↗

The task

When an agent repeats the same website chore every week (read a number, fill a form, download a file), stop paying for it to re-read every page each time.

Steps

  1. You need Python 3.11+: pip install dejo-mcp, then playwright install chromium (about 150 MB; if Chrome is already installed, dejo doctor will use it).
  2. Connect your agent: in Claude Code, claude mcp add --scope user dejo -- dejo-mcp; in Codex, add [mcp_servers.dejo] with command = "dejo-mcp" to ~/.codex/config.toml; in Cursor, add an entry to ~/.cursor/mcp.json.
  3. Run dejo doctor; it names anything missing and the command that fixes it.
  4. The first time, ask the agent for a website task in plain words, e.g. "Go to github.com/microsoft/playwright and tell me how many open issues it has." The agent gets a list of page elements, sends back the steps in one call ending with a read, and Dejo saves the route.
  5. Ask the same thing in a new chat and it is one call. You can also skip the agent and run it from a terminal or cron: dejo run "task name" --site site --json; dejo show site lists the saved steps and dejo forget --site site removes them.
  6. For sites behind a login, dejo_login opens a visible window where you sign in yourself; Dejo keeps the session, never the password.

What the author says

Rohit Kashyap @rohit_jsfreaky (X article, 11 Nov 2026; he builds Dejo): 78% is the average saving on repeat runs and 38% on the first run, both against Playwright MCP on the same 19 tasks. The site describes the method: each task in a fresh Claude session (Claude Sonnet 5, medium effort); Dejo's first run used 251K tokens vs 409K for Playwright MCP, with 19/19 answers right (Playwright MCP 17, Playwright CLI 18); replays used 89K tokens and were a single call on 17 of 19 tasks. In a separate 50-site reliability test, Dejo 0.4.0 found the right element on 47 sites and plain Playwright on 45. He also notes Dejo is not cheaper for single-page, one-line lookups.

Our test

On 12 Nov 2026 on Linux (CPython 3.13.5) we pip-installed dejo-mcp 0.5.1 (Playwright 1.65.1) in a fresh venv; dejo doctor reported python, dejo, browser (chrome), profile, memory and downloads all ok. With a minimal client built on the Python MCP SDK we launched dejo-mcp over stdio and listed 5 tools (dejo_run, dejo_login, dejo_profile, dejo_memory, dejo_manage). The first dejo_run opened https://example.com and returned the element list (2.4 s); sending read e7 as notice saved a 2-step route. Then, from a new process, dejo run "read page heading" --site example.com --json twice returned ok:true with the correct first paragraph, duration_ms 105 and 85, model_calls 0. We removed the route afterwards with dejo forget. We did not connect Claude Code, test logins or site-change repair, or reproduce the author's benchmark.

What we found when we tried it:On a simple page, replays after one learning run needed no model at all: one command, about 0.1 s, with model_calls: 0. Reading by heading text did not save for us; using the element refs Dejo lists was the reliable way.

How to get in

Public worldwide: https://dejo.run and https://github.com/dejohq/dejo-mcp. No account needed.

Who it suits

Developers whose agents repeat the same website chores: weekly numbers, admin forms, post-deploy page checks, price checks. Not useful for a one-off look at a single page.

Pitfalls

  • Replays follow saved steps rather than re-reading the page; when the task itself changes (a different subject), check that Dejo doesn't replay it as the same job. The site says 0.5.1 added checks for this.
  • A learning run must end with a read or nothing is saved; our first attempt with a read by heading text was rejected, and it saved once we used the element ref (e7) Dejo listed.
  • Memory lives locally in ~/.dejo/memory.db, so a new machine starts from scratch; signed-in sessions also live in the local browser profile, so protect it.
  • The 78% and 38% figures are the author's own measurements; we only saw a zero-model-call replay on one simple page.

More in this category