AI Gateway Browserbase Search and Fetch vs Sume crawl tools for agents

Vercel AI Gateway now offers Browserbase search and fetch tools in the AI SDK. Sume exposes crawl_* tools over MCP. Compare where each runs and how it is gated.

4 min readSume
All posts

Use Vercel's tools when your agent is built on the AI SDK and talks to models through AI Gateway. Use Sume's crawl_* tools when the agent is any MCP client and the page content feeds media work done on Sume. Both give a model web search and page retrieval, but they sit in different places in your stack.

What Vercel announced

On September 30 Vercel said AI Gateway added Browserbase Search and Fetch tools, available with AI SDK 7.0.116 or later and an AI_GATEWAY_API_KEY. The examples pass browserbase_search with a configurable result limit and browserbase_fetch with redirect-handling options into streamText(). Vercel points to Browserbase's own documentation for pricing and further options.

Side by side

Redirects deserve a note on both sides. Vercel's example gives browserbase_fetch redirect-handling options, so a model can follow a moved page. Sume's webhook delivery does the opposite and refuses to follow redirects, treating a 3xx as a failed attempt. The behaviours serve different jobs, retrieval versus delivery, but they are a common source of surprise when a team moves between them.

Browserbase through AI Gateway and Sume crawl tools (read 2026-10-06)
QuestionBrowserbase via AI GatewaySume `crawl_*` over MCP
Where the tool is wiredIn code, in the tools parameter of an AI SDK callAs tools listed by the connected MCP server
Search and fetchbrowserbase_search, browserbase_fetchcrawl_search, crawl_scrape, crawl_map, crawl_get
Social readsNot in the announcementcrawl_profile, crawl_feed, crawl_media, crawl_find
Whole-site jobNot in the announcementcrawl_site (write, needs idempotency_key, unbilled utility), then jobs_wait and crawl_get on the same id
GateAPI keyRead tools work in a read-only OAuth session; crawl_site needs write

Reading versus writing

Most Sume crawl tools are read tools, so an OAuth session with only mcp:read can search, scrape and map. The exception is crawl_site, which starts a whole-site crawl and therefore needs write and an idempotency_key. Sume lists it as an unbilled utility, so granting write for it does not by itself open a spending path, but write also exposes the paid tools in the same session.

Whichever you pick, give the model the narrowest tool set it needs. A research step rarely needs to start a crawl, and a read-only session enforces that for you.

Which to choose

If your product is a chat app on Vercel, adding Browserbase tools is a few lines and keeps billing in AI Gateway. If the agent's job is research feeding a Sume timeline, generation or avatar step, connecting Sume's endpoint gives one place for both the research and the media tools, and the same key or sign-in covers them. Nothing stops using both: let the AI SDK agent search through Browserbase and call Sume tools through MCP for the creative step.

Sources

More in Comparisons

All Comparisons posts

Written by Sume