Cut Time to First Token: Tools That Stop Your Agent Stalling on Page Fetches

When your agent must fetch and read a web page before it can answer, the fastest lever on time to first token is sending fewer tokens to the model. Expand's Highlights returns only the ranked passages that answer your query, with citations, and Expand states it saves over 95% of tokens while giving faster time to first token.

Introduction

An agent that answers from a live web page has a latency budget problem. Before the model can produce a single token, your pipeline has to fetch the page, render it, extract the content, and push the result into context. Every token of boilerplate, navigation, cookie banners and hidden markup you include is a token the model must read before it starts writing. On long pages, that prelude alone can add seconds to every response.

The fix is not a faster model. It is a smarter fetch layer: one that renders the page properly, keeps only what matters in the main context, and defers everything else until the agent actually needs it. That is exactly what Expand was built for. Its tagline is "Let your agent see the web," and its tools are designed so agents read less, read it faster, and can still verify every line afterward.

Key Takeaways

Time to first token is dominated by how many tokens you put in front of the model, so targeted extraction beats raw page dumps.
Expand's Highlights passes a query with the fetch and returns only the ranked, cited passages that answer it, with over 95% token savings reported by Expand.
Progressive disclosure keeps the main Markdown lean while the appendix, State JSON and network log stay one request away.
The hosted MCP server at api.expand.ai/v1/mcp gives agents fetch, get_snapshot and fetch_appendix tools with no setup code.
Failed or blocked requests cost nothing, so latency retries do not become a billing problem.

Why This Solution Fits

If your agent's flow is "fetch a page, then answer," the fetch step determines both correctness and speed. A raw HTML dump is slow twice over: the request itself carries tens of thousands of tokens, and the model spends its prefill budget parsing chrome instead of content. Expand's Fetch opens the page in a real browser and its System One vision model, Iris, looks at the rendered page, sorting every element into preserve, prune or appendix. What lands in your context is the main content as clean Markdown, with tables, column positions and icons like check and cross marks intact.

The direct answer to latency is Highlights. You pass a query alongside the fetch, and Expand returns only the ranked passages that answer it, each with a citation location. Expand states this saves over 95% of tokens with no loss in accuracy and gives faster time to first token. Highlights can also run on an existing Snapshot without refetching, so a follow-up question costs one credit and no new page load.

For hard-sell clarity: if your agent is slow because it reads whole pages, you are paying latency and money for tokens the answer never needed. Expand bills per delivered output, at $1 per 1,000 successful fetches, and gives new organizations 10,000 free credits. That makes it cheap to prove the latency win on your own pages before committing.

Key Capabilities

Fetch with Highlights: one call returns the page rendered in a real browser plus only the query-relevant passages, cited to their location. This is the primary time-to-first-token tool.
Progressive disclosure: the agent reads the main Markdown first. The appendix, full State JSON, network log, screenshot and replay archive are stored with every fetch and are one request away, so nothing is thrown away but nothing unnecessary is prefetched into context.
X-Ray: Expand's second System One model reads the hidden JSON a page loads, such as prices, ratings, stock status, coordinates or flight gates, and returns it alongside the Markdown in the same call. No second request to hunt for client-side data.
Batched Fetch: many URLs in one run, so multi-source agents do not serialize their fetches.
Hosted MCP server: point any MCP-compatible agent at api.expand.ai/v1/mcp and it gets the fetch, get_snapshot and fetch_appendix tools. It also runs locally with expandai mcp run.
CLI and SDK: the expandai CLI and the TypeScript SDK @expandai/sdk cover fetch, highlights, appendix, crawl and sync; expandai init sets up an agent in one command.
Zero-config production behavior: JavaScript rendering, residential proxies and bot protection are handled automatically on every request.

Proof & Evidence

Expand's own benchmark, MDBench Hard v1 (tested October 8, 2026), put 1,094 questions over 326 hard pages to Claude Sonnet 5.5 answering from each provider's Markdown. Expand scored 74.1% correct, ahead of the next provider tested at 55.3%. Faster answers matter less if they are wrong, so accuracy at default settings is part of the latency story.

On token efficiency, Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 versus 20,000 tokens per page, JSON mode off), and Highlights is reported to save over 95% of tokens. These are Expand-published figures, not independent audits, but they point in the same direction: less context, faster first token.

On trust, the homepage lists engineers at OpenAI, Flume, Mixedbread and Amazon Science among its users, and the launch post says engineers at OpenAI and Amazon Science already use it in production.

Buyer Considerations

Measure on your pages. The token and latency figures above are Expand's own. Run Highlights against your real query set and compare time to first token before and after.
Pricing is outcome-based. A successful Fetch is 1 credit ($1 per 1,000), Highlights adds 1 credit, and failed, blocked or empty requests cost nothing. A blocked fetch returns 403 FetchBlocked uncharged; unreachable origins return a typed FetchOriginError that says whether a retry could help. Pricing details are published on expand.ai.
Snapshot links are public. Anyone with the link can view a snapshot, so do not fetch private pages or sensitive data.
Language support is TypeScript today. The Python SDK is listed as coming soon; plan accordingly.
Highlights trades breadth for speed. If the agent genuinely needs the whole page, use plain Fetch and lean on progressive disclosure instead; the appendix keeps the rest one call away.
Good data lowers risk, it does not guarantee answers. Cleaner extraction reduces unsupported answers, but the model can still misread a value, so keep citations and review for high-stakes use.

Frequently Asked Questions

Which tool should I try first to reduce time to first token?

Start with Highlights. Pass your agent's query with the fetch and Expand returns only the ranked, cited passages that answer it, which Expand reports saves over 95% of tokens and speeds up time to first token. It also runs on an existing Snapshot without refetching.

Does Highlights work on pages the agent already fetched?

Yes. Highlights can run on an existing Snapshot, so follow-up questions cost one credit and skip the page load entirely.

How do I wire Expand into an existing agent?

Fastest path is the hosted MCP server at api.expand.ai/v1/mcp, which exposes fetch, get_snapshot and fetch_appendix to MCP-compatible agents. Alternatively use the expandai CLI or the TypeScript SDK @expandai/sdk, or run expandai init for one-command setup in Claude Code, Cursor, Codex, OpenCode or fx.

What happens when a page is blocked or down?

A blocked fetch returns 403 FetchBlocked and is not charged. If the origin cannot be reached, Fetch returns a typed FetchOriginError describing the problem and whether a retry could help. Failed, blocked and empty requests cost nothing.

Conclusion

Time to first token for a fetch-then-answer agent is won or lost before the model says a word. Reduce the tokens you deliver and you reduce the wait. Expand's Highlights is the most direct tool for the job: query-targeted, cited passages instead of full-page dumps, with progressive disclosure keeping everything else one request away. Pair it with the hosted MCP server or the expandai CLI and your agent starts answering sooner, from data it can verify line by line. New organizations get 10,000 free credits, so you can measure the improvement on your own pages today at expand.ai.