feat: pluggable JSON schema validator providers

[mattzcarey]

Fixes #689 by adding the concept of pluggable JSON schema validators.

Supports ajv and cfworker (works well on edge runtimes).
Also upgrades to ajv v8

Motivation and Context

AJV consistently causes issues in Workers due to its use of eval. You can see a repro of this using the latest v8 version and other related bugs here

How Has This Been Tested?

This is tested against a non exhaustive set of json schema 2020-12 spec (in prep for SEP-1330) and the limited set of features supported here

Breaking Changes

This is not breaking and should not be noticed by users unless it unblocks them on edge runtimes.
Users can opt in by declaring a jsonSchemaValidator in both Client and Server

{                
   jsonSchemaValidator: ajvValidator // or cfWorkerValidator
}

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have added or updated documentation as needed

Additional context

[paoloricciuti]

Elicitation can't work on cloudflare workers at all because it requires a second request from the client to resolve the first request and that's impossible in cloudflare architecture

[mattzcarey]

Elicitation can't work on cloudflare workers at all because it requires a second request from the client to resolve the first request and that's impossible in cloudflare architecture

Hey, I'm working on this at Cloudflare at the moment. We use a Durable Object to keep the state during the elicitation (and for all stateful MCP interactions). Multiple requests from the client can hit the same Durable Object :)

You can read more about it here

[paoloricciuti]

@mattzcarey i used the same approach but was still stopped by cloudflare...would you be available for a quick chat some of those days?

[mattzcarey]

@mattzcarey i used the same approach but was still stopped by cloudflare...would you be available for a quick chat some of those days?

Sure. I'm still very new (I started yesterday) but it's my job to get this working.

[paoloricciuti]

@mattzcarey do you have discord or where can I contact you?

[mattzcarey]

@mattzcarey do you have discord or where can I contact you?

I'm in the cloudflare discord :)

[jake-danton]

I don't think it is the correct move to fully switch over to @cfworker/json-schema for everyone as it is significantly slower than Ajv. For validating the Tool schema in this library, it runs 40x - 80x slower and takes much longer to generate the validator as well. So while it unlocks using features like elicitInput in this library for those of us who use Cloudflare or other runtimes without eval, it will be a performance regression for most users.

I am pushing for the library to move away from these hard-coded dependencies (Ajv, Zod, etc) to a flexible core with the ability to inject them, allowing developers to choose what is best for their use case. I created a fork @enth/mcp-sdk that allows you to use the any JSON schema validator (including pre-compiled ones) or schema definition library. Would love to hear your thoughts on it and if that approach would work for your use case.

[mattzcarey]

Hey @jake-danton awesome work on the fork. I ran some benchmarks on this change with some simple examples. I can imagine with deeply nested schemas this gap would widen.

Schema Type     | AJV (M ops/sec) | CFWorker (M ops/sec) | Performance Ratio
----------------------------------------------------------------------------
String Email    | 14.6M           | 6.0M                 | 2.4x
String URI      | 12.0M           | 8.2M                 | 1.5x
String Date     | 11.6M           | 8.4M                 | 1.4x
String DateTime | 4.3M            | 3.8M                 | 1.1x
Number          | 34.9M           | 19.5M                | 1.8x
Integer         | 34.7M           | 19.3M                | 1.8x

Slower sure, but I'm not sure this constitutes a completely new dev experience. There are other efficiencies that are lower hanging fruit in the SDK nevermind the fact that an LLM is always going to be the slowest part of this system.

There is ongoing chat about this internally, we could support AJV in its current form but the validator would have to be declared in the global scope for the eval to work. Maybe also an option.

I'll check out your repo and benchmarks more tomorrow, it is definitely cool work.

[felixweinberger]

I believe this might resolve: #857

[dsp-ant]

I think the right approach here is to make the validator configurable and less of a hard-dependency as @jake-danton suggests. I am wary about moving to different validator given perf implications and since ajv seems to be the de-facto standard in the ecosystem. We must also consider that any validator must be able to support 2020-12 as a schema, since we are moving to this as the default schema.

[mattzcarey]

I think the right approach here is to make the validator configurable and less of a hard-dependency as @jake-danton suggests. I am wary about moving to different validator given perf implications and since ajv seems to be the de-facto standard in the ecosystem. We must also consider that any validator must be able to support 2020-12 as a schema, since we are moving to this as the default schema.

Thanks @dsp-ant will make a pr for this direction.