Dejo vs agent-browser
The short answer: agent-browser, from Vercel Labs, is "a native Rust CLI that gives an AI agent a browser it can drive one command at a time": snapshot the page, act on a ref like @e1, snapshot again. Dejo is an MCP server that remembers website tasks: your agent drives it step by step the first time, and every run after that is one tool call with zero model calls. For fast one-off driving from the shell, agent-browser is a strong pick. For website work that comes back, Dejo saves the repeat.
Side by side
| Dejo | agent-browser | |
|---|---|---|
| what it is | MCP server (Python) | command-line tool (Rust), with an MCP mode (agent-browser mcp) |
| how the agent drives it | step lines in dejo_run, many per call | one shell command per step |
| element refs | e12 from look | @e1 from snapshot |
| remembers a task between sessions | yes, replays with zero model calls | not in its README (it keeps the browser and session state) |
| checks each step | yes, every saved step carries a check that it landed | a click fails when the target is hidden, disabled, off screen or covered |
| when a site changes | stops at the one step that moved; your agent fixes it | the agent snapshots and works it out again |
| browsers | Chrome or Chromium, or a Chrome you point it at with DEJO_CDP_URL | Chrome for Testing, a running Chrome, Lightpanda, Obscura, iOS Simulator Safari, cloud providers |
| license | MIT | Apache-2.0 |
| tokens, our real-site benchmark | 251K first run, 89K after | not measured yet |
The agent-browser column comes from its README, read 13 November 2026. agent-browser has not been through our benchmark, so there is no token number for it.
One command at a time, or one call for the whole task
agent-browser is built to make each step cheap. A background daemon keeps the browser open between commands, snapshots are compact text, and refs let the agent act without re-reading the whole page. Its README says to take a new snapshot after a page changes and act on the fresh refs. That loop is quick, and the agent still decides every step on every run.
Dejo cuts the loop out once a task is learned:
- The first run, your agent explores through
dejo_run, which can take many step lines in one call, and Dejo saves each step with a check that proves it worked. - Every run after, one
dejo_runcall walks the whole task in plain code. No snapshots, no model call inside Dejo. - When the site changes, only the step that moved fails, and your agent fixes that one step.
On our benchmark of 19 tasks on 11 real websites, Dejo's runs after the first averaged 89K tokens against 409K for Playwright MCP, and 17 of the 19 tasks replayed in a single call.
When agent-browser is the better choice
- Your agent lives in the shell and you want each browser step to be one short command.
- You need other engines or devices: Lightpanda, the iOS Simulator's Safari, or a cloud browser provider.
- You run agents on Vercel. Its docs cover Vercel Sandbox.
- The task is one-off. There is nothing to remember.
When Dejo is the better choice
- The same website work comes back and you want it done in one call, not one command per step.
- You want proof each step worked on every replay, not only on the first run.
- You want sign-ins kept per profile, with several accounts side by side. See Signing in.
Set up Dejo
pip install dejo-mcp
playwright install chromium
claude mcp add --scope user dejo -- dejo-mcp
Then ask your agent for something on a website, in your own words. Other agents: Codex, Cursor.
More comparisons: Dejo vs Browser MCP, Playwright MCP alternative and Playwright MCP vs Playwright CLI.
FAQ
What is the difference between agent-browser and Dejo?
agent-browser is Vercel's command-line browser for AI agents: the agent runs one short command per step, reads a compact snapshot and acts on refs like @e1. Dejo is an MCP server that remembers each website task, so after the first run the whole task is one tool call with zero model calls.
Does agent-browser have an MCP server?
Yes. agent-browser mcp starts an MCP server over stdio. It also ships skills for agent harnesses, and most people use it as a shell command.
Does agent-browser remember tasks?
Its README describes a browser that stays open between commands and saved session state, not a saved task. Each new task is driven step by step by the agent. Dejo saves a learned task with a check on every step and replays it in one call.
Can I use agent-browser and Dejo together?
Yes. One is a CLI and the other an MCP server. agent-browser suits fast one-off driving from the shell; Dejo suits website tasks your agent repeats.