fix(ui): restore MCP catalog provider logos

[yuneng-berri]

TLDR

Problem this solves:

• Seven MCP catalog logos fall back to provider initials

How it solves it:

• Replace broken CDN URLs with shipped or official logos
• Test actual catalog entries with and without a subpath

User Flow

Before: seven providers show initials instead of logos

1. Open https://litellm-domain/litellm/ui/mcp-servers/
2. Click + Add New MCP Server
3. Exa, Tavily, Slack, Twilio, Playwright, Browserbase, and AWS show initials

After: all seven providers display their logos

1. Open https://litellm-domain/litellm/ui/mcp-servers/
2. Click + Add New MCP Server
3. Exa, Tavily, Slack, Twilio, Playwright, Browserbase, and AWS display logos

Linear ticket

Resolves LIT-7390

Pre-Submission checklist

• Meaningful regression tests added
• Focused local tests pass: 171 tests across eight files
• All 33 required GitHub checks pass; additional CircleCI checks noted below
• Scope is isolated to this issue
• Greptile independently reports 5/5 for the current head

Screenshots / Proof of Fix

Real browser, local proxy and seeded database, SERVER_ROOT_PATH=/litellm; no mocked catalog or images. The local development launcher needed a routing-only adaptation outside the repository to serve the subpath. Screenshots are attached to LIT-7390 and require Linear access. Independent browser QA also passed with screenshots for the same base and head

Before (ae6a4a2)

1. Open http://127.0.0.1:65123/litellm/ui/mcp-servers/ and click + Add New MCP Server
2. All seven affected images have naturalWidth === 0; GitHub and Notion have nonzero widths
3. See attachment MCP catalog logos before ae6a4a2 under /litellm on the linked issue

After (f5bf2da)

1. Open http://127.0.0.1:64345/litellm/ui/mcp-servers/ and click + Add New MCP Server
2. All seven affected images have nonzero widths; GitHub and Notion still load. The five shipped images resolve under /litellm/ui/assets/logos/
3. See attachment MCP catalog logos after f5bf2da under /litellm on the linked issue

Supporting checks

• Default-root browser check at http://127.0.0.1:65324/ui/mcp-servers/: all seven logos load; Search filtering and Exa selection retain the expected name and URL
• Eight focused Vitest files: 171 tests passed, covering catalog filtering/selection, logo resolution, cards, logo selection, create payloads, and chat MCP consumers
• New catalog cases failed before the fix and pass with both empty and /litellm roots
• Restoring the broken Exa URL makes both cases fail; removing the local Exa image also makes both fail
• All five local assets exist in the committed packaged dashboard and return HTTP 200 under /litellm
• Full dashboard ESLint: zero errors; all nine lint budgets pass. Changed-file formatting passes
• Full TypeScript check: 1,841 diagnostics, byte-identical to untouched staging; none added by this change
• CircleCI subpath browser suite: 35 tests passed; UI build and all 33 required GitHub checks passed
• CircleCI OpenAI endpoint test: test_anthropic_with_responses_api sends previous_response_id="hi" and receives 403. Replaying the identical request on merge base ae6a4a2f gives the identical 403; this comes from fix(proxy): authorize every Responses API id, not only the ones the proxy issued #39548, already in staging
• CircleCI default-root browser suite passed: 152 passed, one flaky test passed on retry (keyBudgetWindow.spec.ts). Windows checks passed. Final check summary: 127 passed, one skipped, one inherited endpoint-test failure

Type

Bug Fix

Caveats

Low

• Playwright and Browserbase logos require their official hosts
• Standalone development servers lack the existing /ui asset routes
• One additional endpoint CI failure reproduces on unchanged staging

Final Attestation

• Regression cases and adjacent behavior were exercised as described above

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_mcp-catalog-logos (f5bf2da) with litellm_internal_staging (09b6948)}

[greptile-apps]

Greptile Summary

This PR restores seven MCP catalog provider logos by replacing unavailable CDN URLs and adds registry-driven regression coverage for root and subpath deployments

• Five providers now use assets shipped with the dashboard
• Playwright and Browserbase use official external image sources
• Integration coverage verifies catalog entries, URL resolution, and local asset presence under empty and /litellm roots

Confidence Score: 5/5

The PR appears safe to merge, with the changed logo paths present and correctly resolved across supported server roots

No actionable failures remain; the packaged assets, runtime path handling, test discovery, and production rendering path align with the intended fix

Important Files Changed

Filename	Overview
litellm/proxy/mcp_registry.json	Replaces seven broken provider icon URLs with available packaged or official logo sources
ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mcp_discovery.integration.test.tsx	Adds registry-driven logo regression coverage and moves the component test into the configured integration suite

_{Reviews (1): Last reviewed commit: "fix(ui): restore MCP catalog provider lo..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[devin-ai-integration]

QA passed: real browser, proxy-served UI, before ae6a4a2 vs after f5bf2da. All seven logos load at root and under /litellm. Session

Proof

Before (ae6a4a2)	After (f5bf2da)

Steps: http://localhost:4000/ui/mcp-servers/, click "+ Add New MCP Server", scroll to Search, Web & Browser, Cloud

naturalWidth before -> after: Exa 0 -> 1024, Tavily 0 -> 1024, Slack 0 -> 150, Twilio 0 -> 150, Playwright 0 -> 400, Browserbase 0 -> 200, AWS 0 -> 200. GitHub and Notion 150 on both. Selecting Exa prefills name exa, Streamable HTTP, https://mcp.exa.ai/mcp

Note: on the standalone npm run dev server (port 3000) the five /ui/assets/logos/ images do not resolve because that server has no /ui route. Pre-existing for every /ui/assets reference in the dashboard, not introduced here