Claude 4 API: Tool Use Streaming & Prompt Caching vs Graphlit 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: managed document ingestion and vector retrieval exposed as MCP tools — no pipeline to wire, no chunking strategy to bikeshed, no embedding model to pick. The DX bet is that the right abstraction level is the tool call, not the SDK, and for agent workflows that's actually correct. The moment of truth is registering the MCP server with Claude Desktop and asking it a question about a PDF you just pointed it at — that should work in under 5 minutes and from what I can see, it does. The weekend alternative is Chroma plus LlamaIndex plus a couple Lambda functions, which is genuinely annoying to maintain at scale, so Graphlit earns its keep. What earns the ship is that the tool boundary is clean: you're not adopting a new mental model, you're adding a capabilities endpoint. What I'd flag is the pricing jump from free to $299/mo Pro is steep with nothing obvious in between for serious indie use.”

“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 managed RAG-as-a-service with MCP bindings, and the direct competitors are Unstructured.io for ingestion, Ragie for the managed retrieval layer, and a dozen LlamaIndex Cloud competitors. Graphlit's specific bet is that MCP standardization becomes the default agent tool interface — which is a real bet, not a vague one, and it's pointed in the right direction given Anthropic's push on MCP adoption. The scenario where this breaks is multi-tenant enterprise: when a customer has 500k documents, strict data residency requirements, and needs sub-200ms retrieval, the 'managed service' abstraction starts leaking badly. What kills this in 12 months is not a competitor but OpenAI or Anthropic shipping native file retrieval tools that are good enough for 80% of use cases directly in the API — and that clock is already ticking. What would make me more confident is published latency benchmarks on real document corpora and a credible answer to the data residency question.”

“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 Graphlit is betting on: within two years, agent tool interfaces become the primary way software consumes unstructured data, and MCP wins the protocol war over proprietary agent SDKs. That's falsifiable — if LangChain's tooling or OpenAI's function-calling conventions dominate instead, Graphlit is stranded on the wrong standard. The second-order effect that matters here isn't faster RAG — it's that MCP-native document intelligence commoditizes the retrieval layer and shifts competitive differentiation to the quality of tool orchestration and routing logic above it. Graphlit is riding the MCP adoption curve, and right now it's early-to-on-time: MCP is real but not yet the default. The future state where this is infrastructure looks like: every enterprise AI agent has Graphlit (or something exactly like it) as its document memory layer, the same way every app has an S3 bucket. The dependency that has to hold is MCP becoming a cross-vendor standard rather than an Anthropic-specific pattern — and that's genuinely uncertain.”

“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 here is unclear in a way that's a real problem: is this developer tooling expensed to an engineering budget, or a platform capability sold to an AI team lead? That distinction matters because the sales motion, the pricing anchor, and the competitive set are completely different. The pricing architecture has a structural flaw — $49/mo Starter to $299/mo Pro is a 6x jump with no intermediate tier, which means growth-stage customers churn before they convert rather than expanding. The moat question is the hard one: the ingestion connectors and chunking logic are differentiators today, but Anthropic ships MCP-native file tools, those connectors become table stakes and Graphlit is left competing on managed infrastructure margins, which is not a great business. What would make this a ship is a clear enterprise wedge with a workflow lock-in story — if Graphlit becomes the system of record for an agent's document memory rather than a swappable retrieval endpoint, there's a real business. Right now it reads like a technically sound service with no defensible expansion path.”