MCP resources/read: binary (blob) resources returned as TextResourceContents — base64 string forwarded as text

[juancarlosm]

Version: v1.92.0-dev.2 (same code path on main).

Summary

Reading a binary MCP resource through the proxy returns TextResourceContents (base64 in text, mimeType e.g. application/gzip) instead of BlobResourceContents (base64 in blob). The bytes are intact but mislabeled: a client that reads blob for binary resources sees it empty, and one that treats text as UTF-8 mangles it.

Root cause

_normalize_resource_contents forwards a blob's base64 string as content:

elif isinstance(content, BlobResourceContents):
    normalized.append(ReadResourceContents(content=content.blob, ...))  # content.blob is a str

The MCP SDK then re-serializes ReadResourceContents via create_content (mcp/server/lowlevel/server.py), where ReadResourceContents.content: str | bytes is the discriminator: str -> TextResourceContents, bytes -> BlobResourceContents. Handed a str, it emits text. (The sibling TextResourceContents branch correctly passes content.text, a str.)

Reproduce

1. Upstream MCP server exposing a binary resource (e.g. an application/gzip blob).
2. resources/read through the proxy.
3. Result is TextResourceContents with the base64 in text; blob is absent. Expected BlobResourceContents.

Expected

Binary resources round-trip as BlobResourceContents (base64 in blob), preserving the upstream text/blob distinction.

Suggested fix

Decode the blob to bytes so the SDK re-emits a blob (symmetric with the text branch; the SDK re-base64-encodes, so it round-trips exactly):

elif isinstance(content, BlobResourceContents):
    normalized.append(ReadResourceContents(content=base64.b64decode(content.blob), ...))

[Jupiter363]

I opened #32354 with the narrow fix proposed here

The implementation keeps TextResourceContents unchanged and only changes the blob branch in _normalize_resource_contents: BlobResourceContents.blob is base64-decoded to bytes before constructing ReadResourceContents. That preserves the str vs bytes discriminator the MCP lowlevel server uses when it serializes the response back to TextResourceContents or BlobResourceContents

I also added a regression test for a gzip blob payload and updated the existing blob metadata test to assert decoded bytes instead of the old base64 string behavior

Validation on the PR is green, including the targeted MCP server unit test file, lint, Greptile 5/5, Codecov patch coverage, and CodSpeed