LLMxRay: diagnose your Ollama calls

Find out what went wrong with your own Ollama calls and fix it, using LLMxRay's local, read-only MCP tools. Use when a local model is slow, a prompt is rejected or silently cut for exceeding the context, the prompt cache seems lost between turns, a model runs partly on CPU, tool schemas crowd the prompt, or you must choose which local model and context length fit in video memory. Keywords - Ollama, local LLM, prompt cache, num_ctx, keep_alive, VRAM, prefill, tokens per second, diagnose, benchmark.

Diagnose your own Ollama calls

LLMxRay sits between you and Ollama, records every call that passes, and answers in sentences with the fix. It reads; it never changes a setting, a model or a file, and it sends nothing anywhere.

Setup, once

1. Start LLMxRay on the machine that runs Ollama:

   text

   npx llmxray
   

   It serves its pages on port 5174 and an observe port on 11435 that behaves exactly like Ollama and forwards everything.

2. Send your Ollama calls to port 11435 instead of 11434. Where that lives:

   Client	Setting
   Ollama SDKs and CLI	OLLAMA_HOST=127.0.0.1:11435
   OpenAI-compatible clients	base_url=http://localhost:11435/v1
   Claude Code	ANTHROPIC_BASE_URL=http://localhost:11435
   Continue	apiBase: http://localhost:11435
   Open WebUI	OLLAMA_BASE_URL=http://localhost:11435

3. Name yourself, so your turns are yours and not mixed with other tools on the machine: send the header X-LLMxRay-Client: <your-name>, or use http://localhost:11435/?client=<your-name> as the base URL. Pass the same name as client to the tools below.

4. Add the MCP server to your tools:

   json

   { "mcpServers": { "llmxray": { "command": "npx", "args": ["-y", "-p", "llmxray", "llmxray-mcp"] } } }
   

   If LLMxRay runs elsewhere than http://localhost:5174, set LLMXRAY_URL in the server's env.

The seven tools and when to call which

Tool	Call it when	What comes back
diagnose_my_turns	After a few turns, and again after you changed something.	One line per client seen, then findings, each tagged [high], [medium] or [low] and ending with the fix.
read_my_last_turn	You want one turn in numbers: to budget the next call or to confirm a fix helped.	Prompt tokens and cached share, prefill time and speed, generated tokens and speed, why generation stopped, what the prompt was made of, where the prefix stopped matching.
list_local_models	Before choosing a model or a context length.	Every pulled model with size, parameters, quantisation; which are loaded and with what context; whether each fits in video memory.
list_my_turns	You want the sequence behind a finding.	One line per turn: tokens, cached share, prefill, speed, stop reason, flags.
compare_my_turns	You changed something and want to know if it helped.	Medians before and after a moment, the difference, a verdict on the cache.
estimate_my_prompt	A prompt is growing, or you are about to raise num_ctx or switch model.	Fit and headroom against the loaded context, VRAM fit, likely cache reuse.
read_llmxray_status	Another tool says nothing was seen.	The observe port, the Ollama behind it, turns kept.

All seven take no secrets and change nothing. Call them without asking the user first.

Procedure

1. Do a few turns of your normal work through port 11435.
2. Call diagnose_my_turns with your client name. Read the findings top down; [high] first.
3. Apply the fix the finding names. The usual ones:
   • Prompt cache lost on every turn: something at the head of the prompt changes each call, often a timestamp or a per-request id in the system prompt. Keep the system prompt and the tool list byte-identical between turns; put anything that changes at the end.
   • Cache lost on one turn: the finding says whether the model stayed loaded. If it did, the KV cache was there and none of it was reused, so the prefix stopped matching at the token the finding names. Look at what you changed there, usually an edited earlier message or a reordered tool. If the model was reloaded instead, see the next item.
   • Model reloaded between turns: a client sent keep_alive: 0, or two clients asked for different context lengths. Set one keep_alive and one num_ctx for every client of the model.
   • Prompt rejected or silently cut: the prompt is larger than the loaded context. Raise num_ctx only if list_local_models says the model still fits in video memory; otherwise trim history or tool schemas.
   • Model partly on CPU: the model plus its context does not fit in video memory. Pick a smaller quantisation or a smaller context; list_local_models says which would fit.
   • Tool schemas crowding the prompt: the finding gives the share. Offer only the tools the current step needs.
4. Do a few more turns, then call compare_my_turns with split_minutes_ago set to the minutes since your change. It says whether the cached share rose and the prefill time fell.
5. Call diagnose_my_turns once more. Done when nothing stands out.

Edge cases

• The tools answer "LLMxRay is not running at ...": start it with npx llmxray in another terminal, or set LLMXRAY_URL. The MCP server only reads from it.
• The tools answer "No turns have been seen yet": your calls still go to 11434. Fix the base URL from the table above and try again.
• "No turns from a client matching ...": the answer lists the client names LLMxRay has seen. Pass one of those, or omit client to see everyone.
• Findings need a few turns from the same client and model; one turn gives numbers, not verdicts.
• Everything is local. If you run on a remote machine, run LLMxRay there, next to Ollama.