Treegress MCP - Open-source MCP tool for browser automation

⚠️ Your agent either gets confused or loses context completely

If you've used Playwright MCP beyond simple demos, you've likely seen it: the agent skips elements, misses available actions, or produces incomplete test flows.

The root cause is simple: standard Playwright MCP gives the LLM an ARIA snapshot, not the full set of interactable DOM elements.

That snapshot is only an abstraction of the UI. Even when elements are visible, the agent receives limited and simplified context, which leads to inconsistent decisions and execution errors. Even when accessibility is implemented well, the agent still lacks enough context to understand what elements do and how to use them correctly within the flow.

Why this happens

A side-by-side look at what the agent receives from each approach for the same snapshot

Default Playwright MCP

Sends ARIA snapshot

Incomplete visibility of the website context

May hide real interactable elements from the agent

Agent struggles to complete the task

# SVG Icon Buttons — anonymous clickable images
- img [ref=e29] [cursor=pointer]
- img [ref=e43] [cursor=pointer]
# Toggle Switch — element absent from tree
- generic [ref=e59]: ❌ Inaccessible — div toggle, no ARIA
# Contenteditable — element absent from tree
- generic [ref=e93]: ❌ Inaccessible — contenteditable div, no role
# Custom Dropdown — no children visible
- generic [ref=e51] [cursor=pointer]: Select...
# CSS Toggle — element absent from tree
- generic [ref=e190]: ❌ Single semantic class

Treegress MCP

Serializes the full DOM tree

Returns all interactable elements

Enriches page context so the agent can interpret elements correctly

Gives the agent a fuller, actionable page model to complete tasks on the first attempt

# SVG Icon Buttons — onclick reveals purpose
- generic [ref=e19] [onclick="alert('edit clicked')"] [cursor=pointer]:
    - img [tag=svg]
- generic [ref=e26] [onclick="alert('add clicked')"] [cursor=pointer]:
    - img [tag=svg]
# Toggle Switch — preserved with handler
- generic [ref=e42] [onclick="this.classList.toggle('on')"] [cursor=pointer]:
    - generic
# Contenteditable — attribute surfaced
- generic [ref=e63] [contenteditable]
# Custom Dropdown — hidden children exposed
- generic \"Select...\" [ref=e34] [onclick=\"...dropdown-list...\"] [cursor=pointer]:
    - generic [class=dropdown-list] [hidden]:
        - generic [ref=e35] [onclick=\"event.stopPropagation()\"] [cursor=pointer]: Item 1
        - generic [ref=e36] [onclick=\"event.stopPropagation()\"] [cursor=pointer]: Item 2
        - generic [ref=e37] [onclick=\"event.stopPropagation()\"] [cursor=pointer]: Item 3
# CSS Toggle — class and handler revealed
- generic [ref=e126] [class=toggle-switch] [onclick=\"this.classList.toggle('on')\"] [cursor=pointer]:
    - generic

What your agent gets

More useful output for AI-powered browser automation

Treegress MCP Browser gives your AI agent the full DOM, all interactable elements, and context for action — so it can work from a much more actionable representation of the page than a standard ARIA snapshot provides.

Works with your existing setup

You don’t need to change your workflow. Just replace Playwright MCP with Treegress MCP. Please follow the installation steps in README.md file.

→ Go to GitHub

Better visibility into the real page

Fewer missed UI actions, better reliability

Broader and more seamless test scenario generation

Give your AI testing agent the page context it was missing

⚠️ Your agent either gets confused or loses context completely

Why this happens

What we changed

What your agent gets

Try it in a few minutes