[BUG] Supabase MCP missing from web/desktop Code workspace + stream timeouts

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The /bug command isn't available in this environment. Since these are Claude Code platform issues (not codebase bugs), you'll need to report them directly at https://github.com/anthropics/claude-code/issues.

Here's a formatted report you can paste in:

---

Title: Supabase MCP server missing from Code sessions + stream idle timeouts on edge function deploys

Environment:

• Claude Code web and desktop (same hosted cloud workspace)
• Model: Claude Opus 4.7 1M
• Onset: ~April 21, 2026

Issue 1: Supabase MCP server not available in Code sessions

Account-level Supabase connector is healthy and works in Claude.ai (CTO Console), but mcp__supabase__* tools are not available in Code sessions. Reproduced across: refresh tools list, disconnect/reconnect, fresh sessions, both web and desktop clients.

• Supabase project: limbterdjfdcowaroihh
• GitHub repo: Gen-X-Ops/lux100-engine

Issue 2: Stream idle timeouts during edge function file generation

Reproducible on multi-step edge function builds (~45KB TypeScript). Failure point is the file write step after MCP source pull succeeds. Scope reduction and fresh sessions do not resolve.

• Edge function: vdp-fetch
• April 21: 5+ timeouts across sessions
• April 23: 6+ timeouts across sessions
• Total Code environment failures April 21–26: 12 across 5 sessions

Workaround: Deploying via Claude.ai with Supabase MCP directly. 3 deviations from canonical Code deploy path in 5 days.

---

Two separate root causes are likely: MCP workspace binding not propagating to Code sessions (issue 1), and a server-side stream keep-alive gap during heavy file generation (issue 2). Both are worth filing as separate issues if the GitHub form allows it.

What Should Happen?

For Issue 1: When a Claude.ai account-level connector (like
Supabase) is enabled, Claude Code sessions in the same account
should expose those MCP tools via the deferred tool registry.
Currently, mcp__supabase__* tools are absent from Code sessions
despite the connector being healthy at the account level.

For Issue 2: Edge function file generation operations (single
TypeScript file, ~45KB) should complete within standard timeout
windows without server-side stream idle timeouts mid-write.

Error Messages/Logs

Issue 1 — No error output. The mcp__supabase__* tools are simply 
absent from tool_search results in Code sessions. Sample searches 
attempted by Code:

- tool_search "supabase get edge function deploy" → returns 
  mcp__github__* and Monitor only
- tool_search "+supabase" → no matches
- tool_search "supabase edge function" → returns TodoWrite only

Issue 2 — Stream idle timeouts during file write operations. 
No error stack trace available; the connection drops during 
file generation and the session reports "Request timed out" 
without further detail.

Steps to Reproduce

Issue 1 reproduction:

1. Have a Claude.ai account with the Supabase connector enabled

and confirmed working in Claude.ai chat (CTO Console)

1. Open a fresh Claude Code session (web or desktop)
2. Run: tool_search query="supabase"
3. Observe: only mcp__github__ tools returned, no mcp__supabase__

tools available

Issue 2 reproduction:

1. In a Code session with Supabase MCP working (note: this requires

issue 1 to NOT manifest, which has been the blocker)

1. Open an edge function source file ~45KB or larger
2. Request a multi-line surgical edit to that file
3. Observe: file generation step times out mid-write with no error

beyond "Request timed out"

Note: Issues 1 and 2 may be related infrastructure problems but
manifest as distinct user-facing failures. Filing as one report
per the form's "single bug report" preflight requirement, but
they may need to be split into two issues by the triage team.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown exact version, but Code worked reliably for this workflow through approximately 2026-04-20. Issue onset is ~2026-04-21.

Claude Code Version

Unknown — running on web client (claude.ai/code). Version not exposed in web UI as far as I can determine.

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Filing this report from the CTO Console (Claude.ai with Supabase
MCP) because the /bug command in Claude Code returned
"the /bug command isn't available in this environment" — which
itself may be a separate issue worth flagging: users hitting Code
platform problems cannot use the canonical reporting channel.

This issue has been worked around for 5 days by deploying directly
via Claude.ai with the Supabase MCP server, which has worked
flawlessly across three production deploys (vdp-fetch v60 → v62 →
v63 with 0% deviation on row counts). The CTO Console path is
working; the Code path is not.

Workspace context: this is a hosted cloud workspace (web/desktop
both connect to the same remote sandbox at /home/user/lux100-engine),
not a local-CLI install. The MCP attachment is therefore a
workspace-template configuration on Anthropic's side, not
something the user can configure.

Related context: Anthropic's April 23 postmortem
(https://www.anthropic.com/engineering/april-23-postmortem)
acknowledged Code quality regressions in this exact window.
This report may relate to that broader incident or may be
a separate workspace-template issue.