If you shipped a client library in the last three years, there is a good chance a machine wrote it. Stainless and Fern were the two tools most teams reached for to turn an OpenAPI spec into clean, idiomatic SDKs across a dozen languages — the unglamorous plumbing behind the "install our SDK" button on hundreds of developer-docs pages. In 2026, both of them got bought, five months apart, and one of them is being switched off.
Postman acquired Fern on January 8. Then on May 18, Anthropic announced it had acquired Stainless — the startup that has generated every official Anthropic SDK since the earliest days of the API, and whose generator also produced SDKs for OpenAI, Google, Cloudflare, Replicate, and Runway. Reporting put the deal north of $300 million; the terms were not officially disclosed. The part that matters for anyone with a stainless.yml in their repo: Anthropic said it will wind down Stainless's hosted products, including the central SDK generator. You keep the code already generated for you. You lose the machine that regenerates it the next time your API changes.
The thing being consolidated is not SDKs#
It is tempting to file this under "big company buys small tool, sunsets it." But look at what the tool actually became. A modern SDK generator no longer just emits a Python client. It reads your OpenAPI spec and produces, from the same source of truth, a typed SDK and an MCP server — the connector that lets an agent call your API as a set of tools. Speakeasy and liblab both do this today; liblab will even host the MCP endpoint for you.
So the pipeline everyone treated as boring infrastructure quietly turned into something else: the compiler from "an API exists" to "an agent can use it." That is the single most strategic seam in the agent stack right now. Every API that wants to be reachable by Claude, by ChatGPT, by Cursor, has to cross it.
Whoever owns the compiler from OpenAPI to MCP owns the on-ramp by which every API becomes available to agents.
Now re-read the two acquisitions. A model vendor bought one of the two best compilers and is retiring its shared, neutral version. An API platform bought the other. The layer that used to be Switzerland — a spec goes in, SDKs come out, nobody's allegiance required — now has landlords on both sides.
What that leaves you#
The consolidation is exactly why the remaining independent options suddenly matter more than their star counts suggest. Your realistic replacements, if you're leaving Stainless:
- Speakeasy — the closest like-for-like: multi-language SDKs plus a built-in MCP server generated from your OpenAPI document. It maintains a candid head-to-head comparison of the field, which tells you something about its confidence.
- liblab — generates SDKs and will generate and host an MCP server straight from an OpenAPI spec or a Postman collection, which is the fastest path from "we have an API" to "an agent can call it" without standing up infrastructure.
- OpenAPI Generator — the open-source, Apache-2.0 floor. No MCP server out of the box and rougher output, but it is community-owned and nobody can acquire it out from under you. In a year defined by acquisitions, that is a feature. (If you'd rather hand-write the server than generate it, that's the FastMCP vs. official SDK decision.)
- Fern, still very much alive under Postman, if you want SDKs plus documentation and are comfortable with a platform vendor owning the pipeline.
The move to make first#
Do not start by picking a vendor. Start by looking at your OpenAPI spec, because in this new arrangement it is the only asset you actually own. The SDKs are downstream artifacts; the MCP server is a downstream artifact; the generator that produces them is now a thing that can be bought and turned off with a blog post. The spec is the source. If it is hand-maintained, drifting from your real routes, or missing the descriptions and examples that make a decent MCP tool, no generator will save you — and switching generators will only surface how thin it always was.
The teams that will shrug off the next acquisition are the ones whose spec is complete, versioned, and generator-agnostic enough that swapping the compiler is a config change, not a rewrite. Stainless being retired is a small fire drill for a specific set of customers. The larger signal is that the on-ramp to agents is now contested territory — and the only position that stays neutral is owning the input, not the tool.
