Is anything wrong with the Ollama calls your client sent through LLMxRay on port 11435? A verdict judged over several turns: a prompt rejected or silently cut by the context, the cache lost on every turn or on one turn, a model partly on CPU, tool schemas crowding the prompt. Four tools read your turns, one question each: diagnose_my_turns, is anything wrong (a verdict over several turns); read_my_last_turn, what exactly happened on the latest turn (one turn in full); list_my_turns, what happened turn by turn (the sequence, one line each); compare_my_turns, did a change help (before against after). Call it after a few turns, and again after a change. For which model or context fits use list_local_models. Returns text. First one line per client seen, then one finding per line tagged [high], [medium] or [low], each ending with the fix. It says so when nothing stands out, and explains how to route through 11435 when no turn was seen. client: your X-LLMxRay-Client name; without it on a shared machine you read every tool's turns. since_minutes: 1 to 10080, minutes since your change. They combine. Needs LLMxRay running (npx llmxray) and says so if not. Under a second, from records already kept. No model is called.

What exactly happened on your latest turn through LLMxRay, in numbers: prompt tokens and cached share, prefill time and speed, generated tokens and speed, stop reason, prompt make-up (system, tools, history, new message), context requested, VRAM fit, and where the prompt stopped matching the previous turn. Four tools read your turns, one question each: diagnose_my_turns, is anything wrong (a verdict over several turns); read_my_last_turn, what exactly happened on the latest turn (one turn in full); list_my_turns, what happened turn by turn (the sequence, one line each); compare_my_turns, did a change help (before against after). Use it to budget the next call. For what is installed and loaded use list_local_models. Returns text, one fact per line. Explains how to route through 11435 when no turn was seen. client: a substring of the label LLMxRay shows, which is the name you set in X-LLMxRay-Client. On a shared machine the latest turn may be another tool's, so pass your own name. Omit it when you are the only caller. Needs LLMxRay running (npx llmxray) and says so if not. Under a second, from records already kept. No model is called.

Which models Ollama has pulled on this machine, with size, parameters and quantisation. Which are loaded now and with what context. How much of each sits in video memory, and whether each pulled model would fit. Use it before choosing a model or a context length. It knows nothing about past calls. For those use diagnose_my_turns or read_my_last_turn. Returns text. A machine line (GPU, video memory, RAM) when LLMxRay is running, then "N models pulled:" and one line per model, sorted by name, ending in LOADED now, fits in VRAM, or larger than VRAM. Takes no arguments, no model name and no filter. Needs Ollama reachable. Under a second, two listing calls. No model is loaded. The answer is a snapshot: a loaded model unloads after its keep_alive, so call it right before deciding.

What happened, turn by turn: the turns your client sent through LLMxRay on port 11435, newest last, one line each: turn number, time, client, model, prompt tokens with cached share, prefill ms, generation speed, stop reason, and flags (reloaded, rejected or cut by the context, error). Four tools read your turns, one question each: diagnose_my_turns, is anything wrong (a verdict over several turns); read_my_last_turn, what exactly happened on the latest turn (one turn in full); list_my_turns, what happened turn by turn (the sequence, one line each); compare_my_turns, did a change help (before against after). Use it to see the sequence behind a finding, or to find when a change took effect before calling compare_my_turns with that many minutes. client: your X-LLMxRay-Client name; without it on a shared machine you read every tool's turns. since_minutes: 1 to 10080. limit: 1 to 100, default 20, the newest kept. The three combine. Needs LLMxRay running (npx llmxray). Under a second. No model is called.

Did a change help? Medians of cached share, prefill time, generation speed and prompt size for the turns before and after a moment, with the difference and a one-line verdict on the cache. Four tools read your turns, one question each: diagnose_my_turns, is anything wrong (a verdict over several turns); read_my_last_turn, what exactly happened on the latest turn (one turn in full); list_my_turns, what happened turn by turn (the sequence, one line each); compare_my_turns, did a change help (before against after). Use it after you changed a system prompt, keep_alive, num_ctx or the tool list. split_minutes_ago: 1 to 10080, default 60: turns older than that are "before", newer are "after"; set it to the minutes since your change. client: your X-LLMxRay-Client name, so another tool's turns do not mix in. Both sides need at least one turn, else it says what to move. Needs LLMxRay running (npx llmxray). Under a second. No model is called.

Before sending a prompt of N tokens: whether it fits the context the model is loaded with and the headroom left for the answer. Also whether the model sits in video memory, and how much of the prompt the cache would reuse, from this client's last turn on that model. Use it when a prompt grows, before raising num_ctx, or before switching model. For what is installed use list_local_models; for what actually happened use read_my_last_turn. model: a pulled model name, "gemma4" or "gemma4:latest"; unknown names get the list of pulled ones. prompt_tokens: 1 to 2000000, your estimate of the prompt size. client: your X-LLMxRay-Client name, so the cache estimate uses your own last turn. Needs Ollama reachable; the cache line needs LLMxRay running (npx llmxray). Under a second. No model is loaded or called.

Where LLMxRay runs and what it holds: the observe port and whether it listens, the Ollama it forwards to and its version, how many turns are kept of the 500 it holds, whether request text is kept on disk, and when it started. Use it first when diagnose_my_turns or read_my_last_turn answers that nothing was seen, or to confirm which Ollama your calls reach. Those tools read the turns; this one reads the server itself. Takes no arguments. Nothing can be changed through MCP; settings live on the Live page. Needs LLMxRay running (npx llmxray); if it is not reachable the answer says so and names the URL tried. Under a second.