It could have been, but you still need to compress and transform OpenAPI, here are some examples:
* Stripe OpenAPI has ~300 endpoints, imagine adding it to 128k context window, you will run out of context
* Most APIs return unix timestamp, which LLMs don't understand and can't convert to human readable format without additional tools
* With OpenAPI you are trying to be extensive (see Stripe API, contains everything), with LLMs you want to provide basic functionality (only transaction related functionality for humans, even that is limited to last month or week)
With MCP, I think authors envisioned long running processes with duplex communication (e.g. notification about progress).
You can obviously achieve all these with OpenAPI, but it requires some engineering effort to adjust for LLMs. And when people feed OpenAPI as it is, they would start complaining about sub-par quality, because it didn't understand the timestamp
willahmad•4mo ago
* Stripe OpenAPI has ~300 endpoints, imagine adding it to 128k context window, you will run out of context
* Most APIs return unix timestamp, which LLMs don't understand and can't convert to human readable format without additional tools
* With OpenAPI you are trying to be extensive (see Stripe API, contains everything), with LLMs you want to provide basic functionality (only transaction related functionality for humans, even that is limited to last month or week)
With MCP, I think authors envisioned long running processes with duplex communication (e.g. notification about progress).
You can obviously achieve all these with OpenAPI, but it requires some engineering effort to adjust for LLMs. And when people feed OpenAPI as it is, they would start complaining about sub-par quality, because it didn't understand the timestamp