Claude 4 API: Tool Use Streaming & Prompt Caching vs Firecrawl MCP Server

“The primitive here is clean: incremental tool-call deltas over SSE, and a cache-control header you attach to prompt segments to pin them server-side. The DX bet is that complexity lives in the HTTP layer, not in a new SDK abstraction — you opt in per-request, no new mental model required. The moment of truth is calling `stream=true` on a tool-use request and watching partial JSON arguments arrive before the model finishes thinking, which actually matters for agent loops where you want to dispatch work early. This is not a weekend-script replacement — implementing correct incremental JSON parsing for partial tool arguments plus a reliable distributed cache with 2M token capacity is a real engineering problem Anthropic has solved for you. The specific decision that earns the ship: cache invalidation is explicit and cache hits are reflected in the usage object, so you can actually measure what you're saving instead of guessing.”

“The primitive here is clean: Firecrawl's scrape, crawl, map, and extract endpoints wrapped as MCP tools with proper JSON schema definitions, so any MCP host can discover and call them without glue code. The DX bet is correct — they put the complexity in the server definition, not in the agent developer's lap. First 10 minutes is adding the server config to your MCP host and calling scrape_url; that actually works. The weekend alternative is real — you could wrap Firecrawl's REST API in a quick MCP server yourself in an afternoon — but the official server handles auth, error formatting, and tool descriptions in ways a quick script won't. The specific decision that earns the ship: they didn't invent a new abstraction, they just exposed existing endpoints correctly.”

“Direct competitors are OpenAI's cached completions and Google's context caching in Gemini 1.5 — both shipping for months — so Anthropic is catching up, not leading. The specific scenario where this breaks: cache hit rates depend entirely on prompt structure, and developers who dynamically compose system prompts (inserting user-specific context at the top) will see near-zero cache utilization and pay full price while assuming they're saving money. The prediction: this feature doesn't get killed — it becomes table stakes infrastructure and Anthropic wins by having the largest cache window (2M vs. competitors' current limits). What would have to be true for me to be wrong: OpenAI ships a 10M token cache window before Anthropic's ecosystem matures, commoditizing the advantage. Still a ship because the streaming tool-use delta is genuinely differentiated — no competitor has clean partial-argument streaming for tool calls yet, and that changes agent loop architecture in ways that matter.”

“Category is MCP data connectors; direct competitors are Browserbase's MCP server, Exa's search MCP, and any of the dozen scraping APIs that have shipped similar wrappers. The scenario where this breaks is multi-step crawls inside an agent loop — Firecrawl's async crawl jobs don't map cleanly to synchronous MCP tool calls, and agents that trigger deep crawls will hit timeout and rate-limit walls fast. The 12-month prediction: Firecrawl wins this specific niche because they own the underlying scraping infrastructure, which is the actual hard part. A wrapper built by a third party gets killed; an official server from the team that runs the crawlers has staying power. What would have to be true for me to be wrong: Anthropic ships a native web browsing primitive into the MCP spec that makes specialized scraping servers redundant.”

“The thesis this bets on: by 2027, the dominant AI application architecture is a persistent agent with a large, stable context (tools, memory, instructions) that gets reused across thousands of user interactions — making context I/O cost the primary unit economics lever, not generation cost. The dependency that has to hold: agents don't collapse back to stateless chatbots, and context windows keep growing faster than per-token prices fall. The second-order effect nobody's talking about: prompt caching at 2M tokens makes it economically viable to give every enterprise user a fully-loaded, role-specific agent context at request time — which shifts competitive differentiation from 'who has the best model' to 'who has the best cached context corpus,' effectively making knowledge curation the new moat. This tool is riding the trend of context-window expansion-as-infrastructure, and it's on-time, not early — but the streaming tool-use primitive is ahead of the curve on agent loop efficiency. The future state where this is infrastructure: every production agentic system has a cache manifest the same way it has a CDN config.”

“The thesis: by 2027, AI agents will treat the live web as a queryable database rather than a place humans browse, and the infrastructure layer enabling that is MCP-connected data primitives — not one-off API integrations. What has to go right is MCP adoption continuing its current trajectory as the de facto agent tool protocol, which is a real dependency but one that looks increasingly likely given Claude, Cursor, and the growing host ecosystem. The second-order effect is interesting: if agents can reliably scrape and structure arbitrary web data on demand, the SEO-optimized web becomes agent-optimized, and the teams that get crawled become the teams with distribution. Firecrawl is riding the MCP standardization trend and is early-to-on-time — the spec is young enough that being an official, well-documented server still confers real positioning advantage. The future state where this is infrastructure: every research and monitoring agent has Firecrawl MCP as a default data source the way every backend has Postgres.”

“The buyer is the engineering team at any company running Claude in production with long system prompts or multi-step agents — this comes out of the AI infrastructure budget, not a new budget line, which means no procurement friction. The pricing architecture is sound: cache reads at ~90% discount means the savings are real and measurable in the first billing cycle, which creates immediate retention — developers who restructure prompts to maximize cache hits are now architecturally coupled to Anthropic's caching implementation. The moat question is the honest one: this is infrastructure that OpenAI and Google will match, so the defensible position isn't the feature itself but the ecosystem of developers who've restructured their codebases around it. What survives a 10x model price drop: the streaming tool-use architecture, because that's about latency, not cost. The specific business decision that makes this viable is pricing cache reads as a separate SKU — it lets Anthropic capture value from high-volume production workloads without losing price-sensitive experimenters.”

“The buyer is a developer building an AI agent who needs live web data and doesn't want to manage a scraping infrastructure; the budget comes from dev tools or AI infrastructure spend. The pricing architecture makes sense — it scales with crawl volume, which correlates directly with value delivered, and the MCP server is a free distribution channel that pulls users into paid tiers. The moat question is the real one: scraping infrastructure is genuinely hard to operate at scale, and Firecrawl has built that over years, so the MCP server is a thin layer on a defensible base. The stress test: if Anthropic or OpenAI ships native browsing deeply enough into their agent frameworks that structured scraping becomes unnecessary, this loses relevance — but that's a multi-year risk, not a 12-month one. The specific business decision that makes this viable: using MCP as a zero-CAC distribution channel to convert agent developers into Firecrawl API subscribers is smart wedge thinking.”