MCP server tools fail to register on Windows: server config visible to agent but tools return "No such tool available"

[cfoltz326]

Summary

When the Claude Agent SDK spawns a worker with an MCP server configured via
mcp_servers={...} (stdio transport), the MCP server's instructions appear in
the agent's system context, but none of the server's tools are registered as
callable. Every tool call returns No such tool available. The agent itself
diagnoses the situation, stating "this session appears to be a standard Claude
Code session rather than one launched through [the runner]."

The same runner code, on the same machine, worked successfully on
2026-04-30 (processed 13 vendors end-to-end). It began failing on 2026-04-30
evening with no code changes between the working and failing runs. Failure has
been 100% reproducible across ~6 attempts since.

Environment

• OS: Windows 11, build 26200.8246
• Python: 3.14
• claude-agent-sdk: tested on both 0.1.71 (current) and 0.1.65 — same
  failure on both
• Node.js: v24.15.0
• npm/npx: 11.12.1
• MCP server: @supabase/mcp-server-supabase@latest (stdio transport, spawned via npx)
• Shell: Windows Command Prompt (cmd.exe), not PowerShell

Repro

Minimal runner config that reproduces the failure:

options = ClaudeAgentOptions(
    system_prompt=system_prompt,
    model="claude-sonnet-4-6",
    permission_mode="acceptEdits",
    allowed_tools=[
        "WebSearch",
        "WebFetch",
        "mcp__supabase__execute_sql",
    ],
    mcp_servers={
        "supabase": {
            "type": "stdio",
            "command": "npx",
            "args": [
                "-y",
                "@supabase/mcp-server-supabase@latest",
                "--project-ref="
            ],
            "env": {
                "SUPABASE_ACCESS_TOKEN": os.environ["SUPABASE_ACCESS_TOKEN"]
            }
        }
    },
)

async for message in query(prompt="Run", options=options):
    ...

Run via python runner.py from CMD with both ANTHROPIC_API_KEY and
SUPABASE_ACCESS_TOKEN set in the environment.

Expected behavior

mcp__supabase__execute_sql is available to the agent and SQL queries execute
against the configured Supabase project. (This is what occurred on 2026-04-30
morning.)

Actual behavior

The agent reports the MCP server's instructions are visible in its context,
but the actual tools are not callable. Every attempt to call
mcp__supabase__execute_sql returns No such tool available. The agent
exits without performing any work.

A representative full agent diagnostic message:

"The Supabase MCP server's guidelines are visible in this session (it's
connected at the interface level), but none of its callable tools are
reachable — every attempt returns No such tool available. ... This session
appears to be a standard Claude Code session rather than one launched
through [the runner]."

Manual verification of MCP server

Running the same npx command directly outside the SDK works as expected — the
server starts and waits on stdin (standard stdio MCP behavior):

npx -y @supabase/mcp-server-supabase@latest --project-ref=
[hangs waiting on stdin, as expected]

This rules out the MCP server itself being broken.

Configurations attempted that did not fix the issue

Attempt	Result
Pin claude-agent-sdk to 0.1.65 (down from 0.1.71)	Same failure
Change allowed_tools from specific name to mcp__supabase__* wildcard	Same failure
Change MCP env block from token-only to dict(os.environ) (full env passthrough)	Same failure
Create .claude/settings.json in working directory	Same failure
Move project out of OneDrive to plain C:\ path	Same failure
Restart machine	Same failure

Hypothesis

The SDK's stdio-mode MCP-injection path on Windows + Python 3.14 appears to
be silently failing to spawn the configured server, while still injecting the
server's instructions into the agent's context. The agent then sees a
"connected" server with no callable tools.

I cannot rule out that this is a recent change in the bundled Claude Code CLI
inside the SDK package rather than the SDK Python code itself.

Impact

This blocks all agentic workflows that depend on stdio MCP servers when run
via the SDK on this environment. Direct Claude Code sessions (not via the SDK)
appear unaffected.

Additional notes

Happy to provide full runner source, full agent transcripts from failed runs,
or run additional diagnostics if useful.