Why MCP won — the protocol wars of 2024-25 — step 1 of 8
The graveyard of tool protocols (2023-2024)
By the time MCP showed up in November 2024, the industry had already tried — and abandoned — several other attempts at "let the model call your tool." This lesson is mostly prose. The goal isn't to write code; it's to understand why one specific shape won and the others didn't, so the next time you see a "new universal protocol for X" announcement you can predict whether it'll stick.
If you skip the history, MCP looks inevitable. It wasn't. Multiple big companies tried to own this layer and failed. The reasons they failed are the reasons MCP succeeded, and they generalize beyond AI.
March 2023: OpenAI Plugins launch
In March 2023, OpenAI launched ChatGPT Plugins. The pitch was: any service can ship a plugin, ChatGPT users can install them like browser extensions, and the model can call the service's API when it's relevant. Expedia for flights. Wolfram for math. Instacart for groceries. Zapier for the long tail.
For about six weeks, this looked like the future. Every B2B SaaS company hired someone to "build the plugin." Articles compared it to the iPhone App Store moment for AI. A small ecosystem formed.
Then it stalled.
April 2024: the store quietly closes
A year later — April 2024 — OpenAI announced they were sunsetting the Plugins store. The replacement was "GPTs" and later "ChatGPT Actions," both of which were arguably less ambitious. The original cross-app "any tool can be called by any model" vision quietly died.
Most plugins had been downloaded by fewer than 100 people. The ones that worked were mostly inside OpenAI's walled garden. The flywheel never started spinning. Why?
The fragmented middle: every framework had its own shape
While OpenAI was running the plugin experiment, every agent framework had its own way to declare "here's a tool the model can call":
- LangChain had
Toolobjects and a custom decorator pattern. - LlamaIndex had its own
FunctionToolshape. - CrewAI rolled its own tool spec.
- AutoGen had a different one.
- OpenAI's own SDK had function-calling JSON schemas, which were the closest thing to a standard — but only worked with OpenAI's models.
- Anthropic's SDK had a similar-but-not-identical tool-use JSON schema, which only worked with Claude.
The result was that a tool author had to write the same integration six different ways for six different audiences. Nobody wanted to do that, so most tools got integrated for one framework and the others went without.
The whole space looked like the early-2000s mobile web, where every phone OS had its own browser and developers had to choose one. That era ended only when WebKit ate everything. Until something analogous happened here, "AI tool integration" was going to stay a mess.
Why this graveyard matters
Three things to take from this:
- Protocols don't win on technical merit; they win on adoption asymmetry. OpenAI Plugins worked fine technically. They lost because nobody outside ChatGPT could use them.
- Fragmentation is the default state of any new ecosystem. You should expect six competing standards before one wins. The question is which one has the right adoption profile.
- The pattern repeats. Every layer of the stack — model APIs, agent frameworks, evals, retrieval — has its own version of this fragmentation phase. Watch for the standardizer.
In the next step we look at what specifically OpenAI Plugins got wrong — three concrete failure modes you can use as a checklist on any future "protocol" launch.
Why MCP won — the protocol wars of 2024-25 — step 1 of 8
The graveyard of tool protocols (2023-2024)
By the time MCP showed up in November 2024, the industry had already tried — and abandoned — several other attempts at "let the model call your tool." This lesson is mostly prose. The goal isn't to write code; it's to understand why one specific shape won and the others didn't, so the next time you see a "new universal protocol for X" announcement you can predict whether it'll stick.
If you skip the history, MCP looks inevitable. It wasn't. Multiple big companies tried to own this layer and failed. The reasons they failed are the reasons MCP succeeded, and they generalize beyond AI.
March 2023: OpenAI Plugins launch
In March 2023, OpenAI launched ChatGPT Plugins. The pitch was: any service can ship a plugin, ChatGPT users can install them like browser extensions, and the model can call the service's API when it's relevant. Expedia for flights. Wolfram for math. Instacart for groceries. Zapier for the long tail.
For about six weeks, this looked like the future. Every B2B SaaS company hired someone to "build the plugin." Articles compared it to the iPhone App Store moment for AI. A small ecosystem formed.
Then it stalled.
April 2024: the store quietly closes
A year later — April 2024 — OpenAI announced they were sunsetting the Plugins store. The replacement was "GPTs" and later "ChatGPT Actions," both of which were arguably less ambitious. The original cross-app "any tool can be called by any model" vision quietly died.
Most plugins had been downloaded by fewer than 100 people. The ones that worked were mostly inside OpenAI's walled garden. The flywheel never started spinning. Why?
The fragmented middle: every framework had its own shape
While OpenAI was running the plugin experiment, every agent framework had its own way to declare "here's a tool the model can call":
- LangChain had
Toolobjects and a custom decorator pattern. - LlamaIndex had its own
FunctionToolshape. - CrewAI rolled its own tool spec.
- AutoGen had a different one.
- OpenAI's own SDK had function-calling JSON schemas, which were the closest thing to a standard — but only worked with OpenAI's models.
- Anthropic's SDK had a similar-but-not-identical tool-use JSON schema, which only worked with Claude.
The result was that a tool author had to write the same integration six different ways for six different audiences. Nobody wanted to do that, so most tools got integrated for one framework and the others went without.
The whole space looked like the early-2000s mobile web, where every phone OS had its own browser and developers had to choose one. That era ended only when WebKit ate everything. Until something analogous happened here, "AI tool integration" was going to stay a mess.
Why this graveyard matters
Three things to take from this:
- Protocols don't win on technical merit; they win on adoption asymmetry. OpenAI Plugins worked fine technically. They lost because nobody outside ChatGPT could use them.
- Fragmentation is the default state of any new ecosystem. You should expect six competing standards before one wins. The question is which one has the right adoption profile.
- The pattern repeats. Every layer of the stack — model APIs, agent frameworks, evals, retrieval — has its own version of this fragmentation phase. Watch for the standardizer.
In the next step we look at what specifically OpenAI Plugins got wrong — three concrete failure modes you can use as a checklist on any future "protocol" launch.