io.github.batchwatch/batchwatch-mcp

Current queue conditions for a model right now vs its recent norm — "is it me or them?". Traffic report, not arrival time. Carries n, confidence, freshness. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.

Which provider+model combinations batchwatch currently has data for, and how answerable each is (n, contributors, confidence). Call this first if you are unsure whether a model is measured. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.

The full measured wait distribution for a model + mode over a window. The raw evidence behind the verdict. Returns absence as absence (never zero) when there is no data. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.

How long is MY job likely to wait? Returns the batch-wait distribution (p50/p90) for the model, with n, confidence and freshness. A distribution, never a single ETA. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.

The verdict: given a model and your own max_wait, does the batch queue currently honour it? Returns the batch p50/p90 wait, the sync comparison if known, the cost saving basis, and a graded confidence + why + freshness. It does not decide for you — it reports whether the queue fits YOUR tolerance. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.

The public "is the queue moving now?" reading for a model (delayed without a key). Lighter than should_i_batch — no max_wait judgement, just the current wait numbers with confidence + freshness. batchwatch measures how long jobs ACTUALLY wait in LLM providers' async batch queues. It reports distributions and current conditions, never a point ETA — a personal history of 3-minute medians once produced a 480-minute job, so "your job finishes at 14:32" is a lie the one time it matters. batchwatch does NOT decide for you: you pass your own max_wait and it says whether the queue currently honours it. Every answer carries n (how many measurements), a graded confidence with a plain-English why, a basis naming the evidence, and freshness (how old the newest measurement is). Without an API key the data is delayed, not absent. Absence is reported as absence, never as zero.