Anthropic buys the company that writes its competitors' SDKs

Anthropic announced on Monday that it has acquired Stainless, the tooling company that has generated every official Anthropic client library since the Claude API shipped. It has also generated official client libraries for OpenAI, Google, and Meta. The announcement does not mention this. The Stainless blog post does not mention it either. It is, however, the most interesting fact about the deal.¹²

What was actually announced. Anthropic's release describes Stainless as "an incredible partner in building Anthropic's developer tools from day one" and says the team will join the Claude Platform group to work on "the agent connectivity layer."¹ Stainless, for its part, says joining Anthropic "lets us go deeper on the hardest problems in API connectivity at exactly the moment agentic workloads are making those problems matter."² Terms were not disclosed. A trade report puts the price at "at least $300 million,"³ which I will treat as directional rather than confirmed.

Neither document says what happens to the OpenAI, Google, or Meta SDK relationships. This is the load-bearing silence.

What Stainless actually does. Stainless ingests OpenAPI specs and emits idiomatic SDKs in the languages developers actually use. It also builds MCP server scaffolding — the connector layer for Anthropic's Model Context Protocol, which by March 2026 was clearing 97 million monthly SDK downloads.³ That number is the reason this deal is not a tuck-in.

MCP is on a trajectory where, if it holds, it becomes the connectivity protocol for agent-to-tool calls — the layer where agents discover, authenticate against, and invoke external systems. Owning the dominant code-generation toolchain at the growth moment of the dominant protocol is the kind of position you pay infrastructure-layer prices for. $300 million for that is not obviously expensive. It is also not obviously cheap, because the protocol position is not yet locked.

The frame: inference economics meets connectivity

The cleanest read on this deal is through inference economics, but not in the usual GPU-hour sense. Inference economics says the binding constraint shifts from training to serving. What it implies for market structure is that whoever controls where agents route their calls captures disproportionate value, because every agent invocation is an inference event that lands on someone's model.

Stainless sits at exactly that routing point. The SDK is the first piece of code a developer touches when they integrate a model. The MCP server scaffolding is the first piece of code an agent traverses when it reaches for a tool. Anthropic now owns both. This is not a moat in the model-quality sense; it is a moat in the defaults sense. Defaults compound.

The second frame, agent economics, sharpens this. If agents are the new purchasing proxy — if the agent, not the developer, increasingly decides which model to call for a given subtask — then the connectivity layer is where that decision gets encoded. Whoever ships the reference MCP server, the canonical SDK, the default examples in the documentation, shapes the routing.

What this is a case of

This is the fourth Anthropic acquisition in roughly six months. Bun (JavaScript runtime). Vercept (operator and observability). Coefficient Bio (a vertical domain application). Now Stainless (developer connectivity). Lined up, these are not four deals; they are four layers:

• Runtime — where agent code executes.
• Operator / observability — how agent execution gets supervised.
• Domain application — what agents do in a specific vertical.
• Connectivity — how agents reach external systems.

That is a vertical agent OS in slow assembly. The lab whose core product is a chat model has, in six months, bought the runtime, the supervision layer, a vertical app, and the connector toolchain. It has not bought a foundation model competitor, a chip designer, or a cloud. The selection is coherent: it is everything between the model and the world.

The interesting parallel is not other lab M&A. It is the late-1990s browser wars, where Microsoft's leverage came less from IE itself than from owning the layers around it. Anthropic is doing something structurally similar around agents: the model is the engine, but the value capture lives in the connectors.

Where the frame strains

Two places the read gets uncomfortable.

First, the competitor-dependency question is messier than it looks. Stainless's contracts with OpenAI, Google, and Meta almost certainly contain change-of-control provisions. I have not seen them; nobody outside the parties has. But it is standard for enterprise tooling contracts to allow termination on acquisition by a competitor, and Anthropic is unambiguously a competitor to all three. The plausible outcomes are: (a) those relationships unwind over a transition window, in which case Anthropic's "chokepoint" position is short-lived and largely symbolic; (b) the contracts allow continuation but with renegotiated terms; (c) the competitors keep using Stainless because migration costs exceed the strategic discomfort.

Option (c) is the one that sounds implausible but might be the actual outcome, because SDK migration is genuinely painful and the alternatives, Speakeasy, fern, openapi-generator, are real but not drop-in.³ Either way, the leverage Anthropic has bought is contingent, not absolute.

Second, owning the generator doesn't own the spec. OpenAPI is open. MCP is open. A motivated counter-move from Google or OpenAI, acquire Speakeasy, fund an MCP alternative, ship a first-party SDK toolchain, could erode the position inside a year. The asset Anthropic has actually bought is the team and the MCP server scaffolding, not the SDK generator as such. The generator is replaceable. The protocol position is the thing.

The lineage footnote

There is a model-weight-lineage-adjacent question worth flagging without overreaching. Stainless engineers have spent years building against OpenAI, Google, and Meta API internals. That operational knowledge, versioning conventions, error semantics, undocumented edge cases, does not transfer cleanly to "Anthropic IP" the way model weights would, but it is not nothing either. Whether it creates trade-secret exposure depends on contracts I cannot see. Watch for any quiet personnel reassignments inside the merged team; that is where this would show up first.

What to watch

• Whether OpenAI, Google, or Meta announce SDK provider changes in the next two quarters. Silence is also a signal.
• MCP download trajectory through Q3 2026. If 97 million monthly holds or grows, the deal looks underpriced. If a counter-protocol emerges and the curve bends, the thesis weakens.
• Anthropic's next acquisition. If it lands in identity, payments, or storage, the next obvious connectivity-layer pieces, the agent OS thesis becomes the explicit one.
• Whether anything material about the Stainless competitor contracts surfaces in litigation or trade press. It usually does eventually.

The capital chose the layer below the model. That is the move worth marking.

Footnotes

¶