A layered checklist for when an AI tool will not install, will not connect, loads the model and then dies halfway. Written for the person who has already tried the obvious things.
Most failures in an AI toolchain fall into one of four layers: install and permissions, network and proxy, model and credentials, workflow orchestration. Debugging in layer order is several times faster than poking at the symptom, because a failure in layer N makes every error message from layer N+1 noise.
#
Layer 1 — install
Will not install : check OS version and CPU architecture (x64 vs arm64) before anything else. A wrong-architecture package produces confusing runtime errors, not install errors. #
Installer completes, app will not start : read logs in order — application log, then system event log, then the launcher output. The launcher usually just echoes the application log. #
Two common silent traps : the app installed into the user profile instead of the system location, and antivirus quietly quarantined a binary. #
Exit criterion for this layer : one minimal command works (--version or a health check) before you plug in any real configuration.
#
Layer 2 — connect
Separate DNS, TLS, proxy, and server 4xx/5xx before touching config.curl -v gives you the handshake result in one line; guessing from inside a client wastes the afternoon. #
Proxy setups : check that the routing rules actually cover the target hostname, that the port is not already in use, and thatNO_PROXY is not accidentally swallowing it. #
MCP specifically : test the server on its own before attaching any client. stdio and HTTP transports fail in completely different ways, and mixing them into one debugging session hides the real cause. #
A detail worth knowing : a client that times out can be healthy — the server may be waiting on stdin that never arrives because the client already exited. Check process lifetime, not just the error string.
#
Layer 3 — model
Local model fails to load : VRAM, quantization format, driver version. Try the smallest quantized build first; if that loads, the problem is capacity, not code. #
API errors : authentication (key, quota, region), routing (base_url typo, wrong API version), timeouts. #
Easy to miss : the same key can expose different model sets per region, and the error is usually a plain 404 that reads like "model does not exist" when the model exists somewhere else.
#
Layer 4 — workflow
Runs halfway then dies : add logging first, then make steps resumable. Idempotent small steps beat a clever retry wrapper. #
Do not refactor while reproducing . A bug you cannot reproduce is not a bug you can fix, and "while I am here" changes destroy your ability to tell which edit mattered. #
One variable per change . If you change three things and it starts working, you learned nothing.
#
What I do not take on
Account trading or sharing, cracking, bypassing software restrictions, scraping private data. No exceptions, no clever workarounds.
#
The one-paragraph version
Installs, then connects, then loads the model, then survives interruption. Get all four green before optimizing anything — most "AI is unstable" reports are a layer-3 problem wearing a layer-4 costume.
I do small, fixed-scope debugging and automation jobs — AI tooling, Python, file formats. First diagnosis is free, async and text-only, no calls. Details: how I work.