You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hello,
I get this error when trying to add a server from the Docker MCP Catalog registry:
Server Error: MCPX_UNKNOWN_UNCLASSIFIED
mcpproxy could not classify this failure. Please file a bug report so we can add a specific code.
Cause: failed to connect: MCP initialize failed during no-auth strategy: transport error: failed to send request: failed to send request: Post "docker://mcp/sqlite": unsupported protocol scheme "docker"
The same happens with any server I choose from that repository.
Looking at the error, the issue is clear: MCPProxy is trying to use the docker:// protocol scheme, but Go's HTTP client doesn't natively handle Docker container URLs. The error unsupported protocol scheme "docker" indicates the transport layer is attempting to make an HTTP request to a Docker URL instead of spawning a Docker container.
This is likely in the pkg/transport or internal/client subsystem where transport strategies are resolved. The no-auth strategy path seems to be falling through to a generic HTTP POST handler instead of recognizing the Docker scheme and launching the container.
Steps to isolate:
Confirm MCPProxy version (mcpproxy --version)
Check if Docker Desktop/daemon is running and accessible
Try a non-Docker MCP server from the catalog to rule out general connectivity issues
Look at transport.go for the Docker scheme handler - it may be missing or not registered in the transport factory
Workaround: If you need immediate access to Docker-based MCP servers, you can run them manually via docker run and connect to the exposed port using http://localhost:<port> instead of the Docker registry URL.
Related code area: The error message pattern suggests the transport layer in internal/transport/http.go or similar - the scheme router likely needs a "docker" case that calls Docker SDK to start the container and get its local address.
The fix is available in the latest release: v0.33.1 (via PR #484).
I tested it end-to-end and it works. One minor inconvenience worth flagging: after adding a server from the Docker MCP Catalog, Docker has to pull the image on first run — and during that window the server is shown as disconnected in the UI. I need to think about how to surface that pull progress better (probably an intermediate image_pulling state with a progress hint).
I will continue working on broader registry integration improvements — there is a follow-up spec being drafted to modernize the registry layer around the canonical registry.modelcontextprotocol.io v0 schema, so multi-package servers (npm / pypi / oci) get a deterministic install path instead of the current per-registry heuristics.
Activity
simonlpaige commentedon May 19, 2026
Looking at the error, the issue is clear: MCPProxy is trying to use the
docker://protocol scheme, but Go's HTTP client doesn't natively handle Docker container URLs. The errorunsupported protocol scheme "docker"indicates the transport layer is attempting to make an HTTP request to a Docker URL instead of spawning a Docker container.This is likely in the
pkg/transportorinternal/clientsubsystem where transport strategies are resolved. Theno-auth strategypath seems to be falling through to a generic HTTP POST handler instead of recognizing the Docker scheme and launching the container.Steps to isolate:
mcpproxy --version)transport.gofor the Docker scheme handler - it may be missing or not registered in the transport factoryWorkaround: If you need immediate access to Docker-based MCP servers, you can run them manually via
docker runand connect to the exposed port usinghttp://localhost:<port>instead of the Docker registry URL.Related code area: The error message pattern suggests the transport layer in
internal/transport/http.goor similar - the scheme router likely needs a"docker"case that calls Docker SDK to start the container and get its local address.fix(registry/483): Docker MCP Catalog → docker run install (no fake d…
Dumbris commentedon May 20, 2026
Thank you for pointing this out!
The fix is available in the latest release: v0.33.1 (via PR #484).
I tested it end-to-end and it works. One minor inconvenience worth flagging: after adding a server from the Docker MCP Catalog, Docker has to pull the image on first run — and during that window the server is shown as disconnected in the UI. I need to think about how to surface that pull progress better (probably an intermediate
image_pullingstate with a progress hint).I will continue working on broader registry integration improvements — there is a follow-up spec being drafted to modernize the registry layer around the canonical
registry.modelcontextprotocol.iov0 schema, so multi-package servers (npm / pypi / oci) get a deterministic install path instead of the current per-registry heuristics.albertovelo commentedon May 20, 2026
@Dumbris confirmed, it's now working for me (on a Mac with Docker Desktop) thank you!
feat(070): registry easy upstream-add — keystone core op + CLI MVP (U…
fix(registries): versioned User-Agent for Pulse + Fleur stdio not rem…
fix(diagnostics): classify stdio exit-before-initialize; surface stde…
fix(registries): stop inventing routes on user-added registry URLs, v…