diff --git a/autoassistant/mcp/server.py b/autoassistant/mcp/server.py
index 1fa4891..9fd7732 100644
--- a/autoassistant/mcp/server.py
+++ b/autoassistant/mcp/server.py
@@ -10,13 +10,32 @@
 into `output/`.
 """
 
+import contextlib
 import io
 import logging
+import os
 import sys
+from pathlib import Path
 
 from mcp.server.fastmcp import FastMCP, Image
 
-from autoassistant.mcp import tools
+
+def _pin_config():
+    """
+    autonerves derives its config directory from the process working directory
+    (`conf.instance` defaults to `os.getcwd()/config`) and autofit reads config
+    at import time. A server launched from a foreign directory — e.g. a chat
+    client spawning `wsl.exe`, which lands in a Windows folder — would otherwise
+    scan unrelated files and crash on import (a `desktop.ini` trips configparser
+    interpolation). Pin the config to this assistant's own `config/` so the
+    server runs from any working directory.
+    """
+    from autonerves import conf
+
+    config_path = Path(__file__).resolve().parents[2] / "config"
+    conf.instance = conf.Config(
+        str(config_path), output_path=str(config_path.parent / "output")
+    )
 
 
 def _route_logging_to_stderr():
@@ -37,6 +56,17 @@ def _route_logging_to_stderr():
                 handler.setStream(sys.stderr)
 
 
+# Importing the tool modules imports autofit, which (a) reads autonerves config
+# at import and (b) lets jax's xla_bridge log its backend probe to stdout — both
+# fatal to the JSON-RPC channel. So, before that import: force CPU (skips the
+# jax backend probe), pin the config so it does not depend on the launch
+# directory, and redirect any residual import-time stdout to stderr. Rebind
+# logging handlers afterwards for anything attached during import.
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+_pin_config()
+with contextlib.redirect_stdout(sys.stderr):
+    from autoassistant.mcp import tools
+
 _route_logging_to_stderr()
 
 mcp = FastMCP("pyauto-results-inspector")
diff --git a/skills/af_inspect_results_mcp.md b/skills/af_inspect_results_mcp.md
index 684a1a0..1cc01f7 100644
--- a/skills/af_inspect_results_mcp.md
+++ b/skills/af_inspect_results_mcp.md
@@ -62,6 +62,27 @@ propagates. Anything the stack needs (`PYTHONPATH` for editable/source checkouts
 `NUMBA_CACHE_DIR`/`MPLCONFIGDIR` in restricted setups) must be declared in the
 config's `env` block.
 
+**Windows (Claude Desktop → WSL):** when the interpreter lives in WSL, launch it
+through `wsl.exe`. Nothing beyond `PYTHONPATH` is required — the server pins its own
+config directory and forces JAX onto CPU before importing autofit, so it does not
+depend on the launch directory or extra environment:
+
+```json
+{
+  "mcpServers": {
+    "pyauto-results-inspector": {
+      "command": "wsl.exe",
+      "args": ["-e", "bash", "-c",
+        "PYTHONPATH=/home/you/autofit_assistant /home/you/venv/bin/python -m autoassistant.mcp"]
+    }
+  }
+}
+```
+
+For the Microsoft Store build of Claude Desktop, the config and the failure log
+(`logs/mcp-server-pyauto-results-inspector.log`) live under
+`%LOCALAPPDATA%\Packages\Claude_*\LocalCache\Roaming\Claude\`, not `%APPDATA%\Claude\`.
+
 **Claude Code**: the repo-root `.mcp.json` registers the same server automatically for
 sessions opened in this repo (marginal there — a Claude Code session can already run
 the aggregator in Python — but it costs nothing and exercises the same wiring).