Extract MCP protocol features in a reliable way

[London-Cat]

For now, protocol implementation is extracted using LLM with the MCP server's README.md as a source of information.

Example: https://www.archestra.ai/mcp-catalog/grafana__mcp-grafana?scroll=130

Here is an extraction script: https://github.com/archestra-ai/website/blob/main/app/app/mcp-catalog/scripts/evaluate-catalog.ts

It would be nice to extract MCP features more reliably, considering the actual code.

[Arxemond777]

Hi,
I've started implementing a static analyzer by detecting MCP protocol features.

Currently, I start from this logic.

Scope: Static regex scan (case-insensitive) over source files with extensions
.ts/.tsx/.js/.mjs/.cjs/.py/.go/.rs/.json/.toml/.yml/.yaml, with file count/size limits.

Tools: ✓
Matches any of: add|register|define.*tool, @tool, tools: [...], tools/list.

Prompts: ✓
Matches: add|register|define.*prompt, prompts: [...], prompts/list.

Resources: ✓
Matches: add|register|define.*resource, resourceTemplates or resourceTemplates: [...],
resources/list or resources/templates/list.

Sampling: ✓ (server-side completions)
Matches: completions or (create|list).*completions.
Note: does not verify client-side sampling callbacks.

Roots: ✓
Matches: roots/list, roots/list_changed, or roots near protocol code.

Logging: ✓
Matches: logMessage, serverCapabilities.logging, notifications/logMessage.

STDIO Transport: ✓
Node: process.stdin|process.stdout, stdio
Python: sys.stdin|sys.stdout
Go: os.Stdin|os.Stdout, bufio.NewReader(os.Stdin), bufio.NewWriter(os.Stdout),
json.NewDecoder(os.Stdin), json.NewEncoder(os.Stdout)

HTTP Transport: ✓
Python: fastapi|flask|aiohttp|uvicorn
Node: express|fastify|koa|hono, http.createServer
Rust: axum|actix|warp
Go: import "net/http", http.ListenAndServe, http.Handle(Func)?
MCP hints: text/event-stream|Streamable|SSE

OAuth2 Auth: ✓
Matches: oauth2|OAuth2, WWW-Authenticate, resource_metadata, AuthorizationCodeBearer|OpenID.

Gradually improving it and comparing my result with the already existing one here https://www.archestra.ai/mcp-catalog?search=mcpmcp-server&sort=relevance&dir=desc and the target repo.

If anyone wants to discuss, I'll be glad to connect.:)

[London-Cat]

@Arxemond77, I wonder why you chose the static analysis over running MCP servers and querying the API? From the example you provided, it's hard to see how it will detect the feature implementation across different languages.

[London-Cat]

@iskhakov @joeyorlando FYI

[Arxemond777]

@London-Cat

1. Why not run?
   I saw in the chat that there was already an attempt to run servers https://archestracommunity.slack.com/archives/C0975QYN5LZ/p1755582479831699 without any success. I'm thinking about trying to extend it, find a similar one, or implement it on my own, in case static analysis won't work
2. How to implement across different programming languages?
   It uses language-specific regex heuristics for each feature -> frameworks, libraries e.g. fastapi/flask/aiohttp for Python, express/fastify/koa for Node, axum/actix/warp for Rust, "net/http", http.ListenAndServe for Go

[Arxemond777]

Hi @London-Cat @iskhakov @joeyorlando

Pls, review my PR for this issue #77

[Arxemond777]

Hi @London-Cat @iskhakov @joeyorlando

I created a new PR