Skip navigation
Playwright CLI + Jev is 98% cheaper and TWICE as fast compared Playwright MCP
0:00 / 0:20

Playwright CLI + Jev is 98% cheaper and TWICE as fast compared Playwright MCP

Filip Hric

@filip_hric3.7K followers

Visit on X
501 likesDownloadOriginal post

42,163 views

Playwright CLI + Jev is 98% cheaper and TWICE as fast compared Playwright MCP Decided to test out Jev by TypeSafe AI and hook it up to Playwright CLI. I compared the results with Playwright MCP to see how both workflows handle a fuzzy instruction such as: "Create new board, list and card" for my Trello app playground. Playwright MCP does what Playwright MCP does - snapshots, decisions thrown to LLM and execution with tool calls (I used Astra). With Jev, I hand it a snapshot from Playwright CLI and let it decide which element to interact with. Probably could be optimized even better, but I’m quite impressed with the results already.

14 Comments

Comments from the original post on X (Twitter)

  • Antoine's profile picture

    @TheAntoineGenes

    If you could run this with a local Laya model or CLM from huggingface, that’d make it 100% free 😊

    2 likes

  • adrian 🤞🏿's profile picture

    @qualityadrian

    i was wondering if this would work.

    1 like

  • Buswe's profile picture

    @buswe_com

    A big chunk of that saving is context: the tool schemas ride along on every turn, while a CLI only pays for the calls it actually makes.

    1 like

  • Nitiksh · NTXM's profile picture

    @Nitikshofficial

    This lines up with what I keep seeing. Thin CLI wrappers around Playwright win on cost and latency for fuzzy browser tasks. Full MCP stacks add flexibility, but the round trips and context bloat show up fast. The sweet spot for me is a headed browser session you keep alive, plus a small action API (map, click, type, wait, screenshot) so steps stay replayable. Fresh profiles every run is where logins die. I wrote up that bring-your-own-session approach here: Curious whether your CLI path kept cookies across the fuzzy task or re-authed each time.

    1 like

  • Ábel Énekes's profile picture

    @enekes_abel

    I also tried it myself, it's definitely making things cheaper for quick, trivial tasks. But quite early it showed that Jev is a non-thinking model so I had to give it an out to fallback to llm when it couldn't complete the goal or had to fill freetext values. Now I'm playing with exposing my Page Object to Jev so it can call their actions directly. Works surprisingly well for complex flows.

    1 like

  • Keith Tsang's profile picture

    @kidtsang

    Jev’s efficiency is impressive, but the real advantage lies in adaptability. Playwright CLI + Jev can handle more complex scenarios with less friction, which is often overlooked. Speed is great, but versatility can be a game changer.

  • Shafie's profile picture

    @shafie_mukhre

    wait, playwright cli can give browser snapsot?

  • Filip Hric's profile picture

    @filip_hric

    Yup

  • Shafie's profile picture

    @shafie_mukhre

    okay I will check again, I was experimenting with a self healing UI testing scripts and decided to go with Playwright MCP because Playwright CLI can’t do browser snapshots (at least not the same as how MCP does it - maybe), but I should actually try

  • GS.ai's profile picture

    @gauravsainicom

    Why not run it locally. Meet 🤝 Kevin

  • Samuel Hu's profile picture

    @realSamHu

    98% cheaper is the kind of result that makes the receipt matter. i’d log the exact Playwright action, latency, and failure branch too. did Jev hit the same happy path or a different retry gate?

  • Filip Hric's profile picture

    @filip_hric

    same paths. Jev actually did more tool calls because it did extra re-checks before filling the input fields and checks browser state before every action, so if that part could be optimized it might slash the execution time even further the mcp flow goes like this: Snapshot → LLM chooses action → MCP executes → updated snapshot → repeat the jev flow goes like this: DOM observation → Jev chooses action → CLI executes → updated observation → repeat

    1 like

  • Samuel Hu's profile picture

    @realSamHu

    extra re-checks before input fill are the agent paying for its own uncertainty. snapshot once, act on the snapshot, and only re-read when the write actually lands.

  • Sujal's profile picture

    @hecodesforme

    the cli snapshot is text so jev never sees the render, two elements with the same accessible name look identical