Web Speed Kills the Token Tax: 90% Cheaper Web Agents via Deterministic DOM-to-JSON
Web Speed launched on Product Hunt today. The pitch is one line — kill the token tax, 90% cheaper agents. The actual product is a logic layer that sits between an agent and a webpage and converts raw DOM into deterministic JSON machine maps. The point is that asking an LLM to read 80KB of nested HTML to find a button is a dumb use of tokens. Pre-parse the page into a structured representation, hand the agent only what it needs, save 70 to 90% of the bill.
Delivery is an MCP server plus a multi-language SDK. Plug into Claude or Gemini, point it at a URL, get back a JSON map that an agent can reason over directly. The deterministic-mapping engine means the same page gives the same JSON every time — no stochastic re-parses, no surprise token blowups. Token budget becomes a planning quantity instead of a runtime risk.
The interesting framing — web agents lose money on cheap pages. The conventional debate is about expensive long-context tasks. Web Speed is pointing at a different leak — every single navigation on a heavyweight page burns tokens for almost no information gain, and at agent scale this is the dominant cost. Strip that out and the unit economics of consumer web agents stop being upside-down.
Place this next to PageIndex (May 8), which did something analogous for documents — reasoning-based retrieval instead of embedding-based — and the broader vectorless-RAG cluster from the past two weeks. Pattern across all of them — strip the LLM out of the parsing path. Use deterministic structure where you can, save the model tokens for the parts that actually need a model. The agent infrastructure stack is increasingly about routing tokens to where they pay off, not about feeding more tokens in.
getwebspeed.io. PH launch today. The 90% number deserves verification on contested pages — SPA-heavy apps, dynamic content, paywalls — and the founders acknowledged those questions in the launch thread. Worth tracking through the next month for production case studies.
← Back to all articles
Delivery is an MCP server plus a multi-language SDK. Plug into Claude or Gemini, point it at a URL, get back a JSON map that an agent can reason over directly. The deterministic-mapping engine means the same page gives the same JSON every time — no stochastic re-parses, no surprise token blowups. Token budget becomes a planning quantity instead of a runtime risk.
The interesting framing — web agents lose money on cheap pages. The conventional debate is about expensive long-context tasks. Web Speed is pointing at a different leak — every single navigation on a heavyweight page burns tokens for almost no information gain, and at agent scale this is the dominant cost. Strip that out and the unit economics of consumer web agents stop being upside-down.
Place this next to PageIndex (May 8), which did something analogous for documents — reasoning-based retrieval instead of embedding-based — and the broader vectorless-RAG cluster from the past two weeks. Pattern across all of them — strip the LLM out of the parsing path. Use deterministic structure where you can, save the model tokens for the parts that actually need a model. The agent infrastructure stack is increasingly about routing tokens to where they pay off, not about feeding more tokens in.
getwebspeed.io. PH launch today. The 90% number deserves verification on contested pages — SPA-heavy apps, dynamic content, paywalls — and the founders acknowledged those questions in the launch thread. Worth tracking through the next month for production case studies.
Comments