“Not natively.”
But you can still ask.
Intervals.icu records more than it draws. Ask the community how to track one specific thing across a season and the honest answer is often that you would need to build a custom field for every variant you want — one for 5-minute VAM, another for 20-minute, another for the next idea you have on a ride.
The data is already there. What follows is a set of questions people keep running into, the prompt that answers each one, and where each answer stops being trustworthy.
A trainer ride that arrived as one long interval.
A session driven in ERG mode and synced from a trainer app usually lands as a single manually-added interval. Interval analysis has nothing to report, even though the power trace is cleaner than anything you record outdoors.
Analyse [activity]. It synced from a trainer app as one interval — find the actual efforts in the power trace and give me each rep's duration, average power, heart rate, cadence, and how much I drifted across it.
What comes back: each steady effort segmented from the trace, with its own statistics and the transitions excluded, so a five-second surge between reps does not drag the average of either.
The limit. Detection reads the power stream, so a ride recorded without a power meter cannot be segmented this way. Where the activity does have real recorded intervals, those are the better source and are read directly instead.
Which climb was steepest, and where was my VAM best.
Elevation is recorded on every outdoor ride and segmented on none of them. Tracking climbing rate across a season means building a custom field per duration, and you have to decide the durations before you know which ones matter.
Find every climb over 100 m in my rides since June. Rank them by VAM, and tell me whether my power held to the top or faded on the longest three.
What comes back: each sustained climb with its length, height gained, average and steepest gradient, duration, VAM, and how power and heart rate behaved from bottom to top. A switchback that gives back a few metres does not split one climb into two, and a stretch of rolling road that merely trends upward is not counted as a climb at all.
The limit. Needs elevation and distance, so an indoor ride has nothing to find. Climbs are detected from the shape of the trace, not matched against a route database — it will describe the climb precisely and will not tell you its name.
The same interval, eight weeks apart.
You repeat a session every week. The whole-activity comparison is too coarse — the warm-up drifting tells you nothing — and a scatter of every interval in the ride puts your Set 1 dots in the same cloud as your cool-down. What you want is one labelled block, tracked.
Compare the "Set 2 sweet spot" interval across my last six sessions. Before reporting anything, check the blocks are comparable: same label, duration within 10%, similar intensity. If they are not, say so and do not report decoupling. List every block you excluded and why.
What comes back: that block alone across the sessions that genuinely match, with power at a given heart rate, heart rate at a given power, and the ones it refused to compare.
The limit, and it is the important one on this page. An assistant will happily compute decoupling on two blocks of different lengths and hand you a number that looks like adaptation. Heart rate lags a change in power by 30–60 seconds, so a longer block reads lower on decoupling whatever your fitness did. The instruction to refuse is doing real work — without it the answer is confident and wrong.
A metric that nobody has built a chart for.
Efficiency factor on threshold intervals only. Intensity factor over a block. A power distribution across a date range. Quadrant analysis. Each of these is a reasonable question and each currently needs its own custom field, built before you know whether the answer is interesting.
Work out efficiency factor for my threshold intervals only, month by month since January. Show which activities you used, which you excluded, and what the trend does or does not support.
What comes back: the calculation, the inputs, and its own working. Your assistant can also reach fields a fixed response shape never returns — time in each power or heart-rate zone, weather, or anything else recorded on the activity — because it can look up what exists rather than guess.
The limit. This is computed when you ask, not maintained on a dashboard. If you want the number in front of you every morning, a custom field is still the right tool. This is for the question you have once, or the one you did not know you had until Tuesday.
Is this number real, or is my power meter lying?
A dropout, a stuck reading, a sensor that spent four minutes reporting the same watt figure. Averages computed over those stretches look exactly like averages computed over good data, which is what makes them dangerous.
Analyse [activity] and tell me whether any of the power data is untrustworthy before you report averages.
What comes back: flagged ranges are reported alongside the figures, and any average or maximum spanning one is marked rather than presented plainly. Where a climb's power is affected but its elevation is not, the warning says so — gradient and VAM stay trustworthy through a power fault.
The limit. Detection is deliberately conservative. It requires cadence to contradict the power reading, because a trainer holding ERG produces long runs of one exact value by design. That means a genuine sprint passes without comment — and so can an isolated meter glitch.
What actually happened in the last twenty minutes.
Summary fields describe an activity. They cannot tell you where the wheels came off, whether you faded before or after the climb, or what your cadence did when the pace dropped.
Pull the per-second power, heart rate, and cadence for the last 20 minutes of [activity] and tell me where the effort changed and what changed first.
What comes back: genuine one-second samples, at whatever resolution the question needs, paged so a four-hour file can be walked completely rather than truncated at an arbitrary point.
The limit. Anything that will not fit in one response is paged rather than dropped, and the response says how much is left. Asking for four hours of raw data in a single answer is a request nothing can satisfy — narrow the window and it comes back whole.
And “this week” meaning this week.
A small thing that ruins larger answers. An assistant asked about recent training will otherwise resolve the date against whatever it assumes today is, which matters most in exactly the question where it is hardest to notice — comparing a block against the same weeks last year.
Every date-range answer carries the real calendar it was resolved against, so “the last eight weeks” means the last eight weeks and a year-on-year comparison lines up the dates it says it did.
Free for 30 days, then free for less.
Every new account gets 30 days of full access. After that, Free keeps working at one pull per tool per day, which covers a real look at your training; Pro removes the limit at $5/month or $49/year. Works with Claude, ChatGPT, Grok, or any other assistant that supports remote MCP — one connection, and nothing to install.