mcp: support http(s) URLs for spec_path in OpenAPI MCP loader

[moophlo]

Relevant issues

Fixes documentation/implementation mismatch where spec_path is documented as supporting URLs, but the loader only supported local filesystem paths.

Type

🐛 Bug Fix

Changes

• Fix spec_path handling in MCP OpenAPI loader to support both:
  • local filesystem paths
  • http(s) URLs
• Align implementation with documented behavior (spec_path: path or URL)
• Add unit test covering OpenAPI spec loading from an HTTP URL
• No behavior change for existing file-based configs

Pre-submission checklist

• ✅ Added tests under tests/mcp_tests/

Test results

Ran MCP unit tests successfully (excluding two tests that are unrelated to this change and currently fail due to import/runtime environment requirements):

poetry run pytest tests/mcp_tests -x -vv -n 4 \
  --ignore=tests/mcp_tests/test_mcp_hooks.py \
  --ignore=tests/mcp_tests/test_proxy_mcp_e2e.py

Result: 89 passed, 5 skipped

Note: Local test run emits asyncio warnings (Task was destroyed but it is pending!) originating from background tasks in MCPServerManager. These warnings are pre-existing and not introduced by this change.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 9, 2026 2:39pm

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
3 out of 4 committers have signed the CLA.

✅ Sameerlite
✅ emerzon
✅ CAFxX
❌ Andrea Odorisio

---

Andrea Odorisio seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Overview

Greptile Summary

This PR aligns the MCP OpenAPI loader’s documented spec_path behavior with implementation by allowing spec_path to be either a local file path or an http(s) URL. It does this by updating load_openapi_spec() to detect URL schemes and fetch the spec over HTTP, and adds a unit test that monkeypatches httpx.get to validate URL loading.

Main integration concern: URL loading currently uses httpx.get directly rather than the project’s configured/custom httpx client path used elsewhere in the MCP tooling, which can cause URL-based specs to fail in environments depending on that configuration (proxy/TLS/auth/instrumentation).

Confidence Score: 4/5

• Mostly safe to merge, but URL spec loading should use the project’s configured HTTP client to avoid deployment-specific failures.
• The change is small and covered by a unit test, but it introduces a separate HTTP code path (httpx.get directly) that bypasses existing httpx client configuration used elsewhere in the module, which is a common source of real-world breakage when proxies/TLS/auth are required.
• litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py

Important Files Changed

Filename	Overview
litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py	Adds URL support to load_openapi_spec() via direct httpx.get; main concern is it bypasses the project’s custom httpx client/config used elsewhere in this module.
tests/mcp_tests/test_openapi_spec_path_url.py	Adds a unit test that monkeypatches httpx.get to validate load_openapi_spec() can load an OpenAPI spec from an http URL.

Sequence Diagram

sequenceDiagram
    participant Caller as MCP OpenAPI Loader
    participant Loader as openapi_to_mcp_generator.load_openapi_spec
    participant URLParse as urllib.parse.urlparse
    participant FS as Local filesystem
    participant Httpx as httpx.get

    Caller->>Loader: load_openapi_spec(spec_path)
    Loader->>URLParse: urlparse(spec_path)
    alt spec_path scheme is http/https
        Loader->>Httpx: GET spec_path (timeout=30s)
        Httpx-->>Loader: Response (json)
        Loader-->>Caller: dict spec
    else local path
        Loader->>FS: open(spec_path)
        FS-->>Loader: file contents
        Loader-->>Caller: json.load(file)
    end
    Caller->>Caller: get_base_url(spec)
    Caller->>Caller: register_tools_from_openapi(spec, base_url)

Loading

[greptile-apps]

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Bypasses configured HTTP client

load_openapi_spec() now fetches URL specs via httpx.get(...) directly, but the rest of this module uses get_async_httpx_client(..., llm_provider=httpxSpecialProvider.MCP) to ensure the project’s custom httpx configuration (proxies, TLS settings, auth headers, instrumentation, etc.) is applied. As-is, loading specs from URLs will ignore those settings and can fail in deployments that require them. Prefer using the same shared/custom httpx client path for URL loads as well.