LLMs spend most of their context window on DOM structure that does not change between turns. Our optimised DOM export strips that overhead with short role codes, single-child flattening, adjacent text merging, and empty-field elision.
Same workload, same input page, same agent prompt. Three snapshots, measured by the GPT-4 tokenizer.
Tokens. Standard accessibility tree, the baseline.
Tokens. 73% reduction vs Chrome AX.
Tokens. 93.1% reduction vs Chrome AX. 74.6% vs Playwright.
We emit a flat node array of [depth, roleCode, text, ...optionalAttrs]. Roles are short codes (doc, btn, lnk, h1); structural wrappers without semantic value are skipped; single-child wrappers collapse into their child; adjacent text nodes merge; empty or default fields drop entirely. The result is still semantic JSON the LLM can navigate, just denser.
One small wire format. Tunable via maxDepth, maxNodes, and maxTextLength params.
{
"v": 1,
"title": "Sign in",
"url": "https://example.com/login",
"nodes": [
[0, "doc", "Sign in"],
[1, "h1", "Welcome back"],
[1, "lbl", "Email"],
[1, "tbox"],
[1, "lbl", "Password"],
[1, "tbox", null, { "type": "password" }],
[1, "btn", "Sign in"]
]
}docs.vulpineos.com/architecture — every patch, where it lives, and why.
github.com/VulpineOS/VulpineOS — runtime, patches, additions.
Pull the source, build the binary, and drive your first agent. The waitlist gates the managed runtime; the open-source path is open now.