Claude 4 API: Tool Use Streaming & Prompt Caching vs Firecrawl MCP Server v2

“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 is clean: MCP-exposed tool that takes a URL and a JSON schema, returns typed structured data. That's the right abstraction — it moves the extraction concern out of the agent's prompt and into a proper typed contract, which is exactly where it belongs. The DX bet is putting schema definition at call-time rather than requiring pre-configured extractors, and that's the correct call for agent workflows where the target schema is determined at runtime. The JS rendering and session cookie support closes the gap on the 'but my target site uses React and auth' objection that kills most scraping tools in real use. The one thing I'd want to verify before fully committing: does the structured extraction degrade gracefully when the schema doesn't match the page, or does it hallucinate field values? That failure mode is the entire ballgame for agents relying on this for downstream logic.”

“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.”

“Direct competitor here is Browserbase plus a schema extraction prompt, or just Playwright with a structured output call to GPT-4o — both are DIY but entirely viable. What Firecrawl v2 actually buys you is the MCP integration layer and the managed rendering infrastructure, which is real value if you're building agents and don't want to operate headless browser fleets. The scenario where this breaks is high-volume scraping of anti-bot-protected sites — Cloudflare and similar will eat through session cookies in ways that require more sophisticated fingerprint rotation than a managed service typically provides. The 12-month kill scenario: Anthropic or OpenAI ships native web retrieval with structured output as a built-in tool call, which is not a crazy bet given the trajectory. What would have to be true for me to be wrong: enterprises get locked into Firecrawl's reliability SLAs and the switching cost becomes real before the platform players close the gap.”

“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 here is falsifiable: by 2027, AI agents will need web data as a typed, structured input — not as retrieved text to be re-parsed — and the tooling layer that provides this will be infrastructure, not a feature. Firecrawl is betting on MCP as the winning protocol for agent tool composition, which is an on-time-to-slightly-late bet given MCP's adoption curve is already steep. The second-order effect that matters: if structured extraction at the MCP layer normalizes, it shifts power from data aggregators (who sell clean datasets) toward agents that can self-serve structured extraction on-demand, which compresses the value of static data products. The dependency that has to hold is MCP remaining the dominant agent tool protocol rather than getting fragmented by competing standards — that's not guaranteed, but it's plausible enough to build on. If this wins, Firecrawl becomes the database driver for the web-as-a-data-source stack.”

“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 or small team building an AI agent that needs reliable web data, and the budget comes from infrastructure spend — that's a real line item with precedent. The pricing architecture is credit-based against usage, which aligns with value delivered and scales with the customer's own growth, but the jump from $83/mo Standard to $333/mo Growth is steep enough that mid-scale users will either cap out awkwardly or overpay. The moat question is the hard one: the technical differentiation is thin against a well-funded competitor who decides to build MCP-native extraction, and 'managed rendering infrastructure' is not a durable moat unless they build proprietary anti-detection capabilities that are genuinely hard to replicate. What makes this viable in the near term is distribution — they have brand recognition in the web scraping space and a developer community that already trusts the API, which is a real head start even if the technical moat is shallow.”