Add Pacevera to Claude Desktop
Install the current Pacevera MCPB in the AI host supported today. No Pacevera account is created.
Adaptive training, made accountable
Know what to do today.
Pacevera turns your training evidence into an explainable change to the workout you already planned. Keep your AI coach. Add a decision layer.
Change the conversation window and you do not have to tell it your training history again. It runs on your own machine, inside the AI you already use.
local-first engine
one file · no API key
reviewed synthetic cases
DC-001, a reviewed synthetic case replayed on this page. Planned: VO₂max Intervals, 60 min · High. Read on the athlete's own computer from Apple Health and Strava. Decided: Moderate run, 45 min · Moderate. Acute:chronic workload ratio 1.58 is above 1.4: load has been ramping too fast. At moderate intensity the session is no longer "VO₂max Intervals"; it becomes "Moderate run". Only 45 minutes are available, so the session is shortened. Confidence medium.
DC-001 · synthetic case · engine 1.12.0 · rules 1.9.1
A reviewed synthetic case, replayed. This page reads nothing from your computer.
One source is enough: Strava alone adjusted one, Garmin alone adjusted one, Oura alone kept one and Whoop alone deferred one. Nobody has to complete the set.
See it in action
Which of these
is today?
Four situations, four replayed cases. Pick the one that sounds like your morning and the worked example below switches to it — same question, same evidence, same decision the engine returned.
You, to your AI
My plan says VO₂max intervals, 60 minutes. I've got 45 minutes today. Should I do it?
Your AI checks today’s Evidence
Pacevera checked the evidence and returned an accountable decision.
Your AI, in plain language
Moderate run, 45 min · moderate. Acute:chronic workload ratio 1.58 is above 1.4: load has been ramping too fast.
You follow up
Why did today’s session change?
You report what happened
These 4 situations replay reviewed synthetic cases through the decision engine (1.12.0 · rule library 1.9.1). Nothing on this page reads your computer.
Five decision types, one engine. Keep is a decision too — checked, and cleared to execute.
How it works
Ask once.
Get a decision you can inspect.
Start with the AI host you already use, choose how today’s Evidence enters your own computer, then ask with the session already on your calendar.
Install Pacevera in the currently supported AI host. Choose local import, spoken Evidence, or a separately authorized Google Health connection. Ask about today’s planned session to see the decision, change, and reason.
Use the entry that fits today
Each is a separate way in; no one needs to connect every source.
Local import · Garmin, Apple Health, or Strava exportSpoken Evidence · tell your AI what happenedGoogle Health · authorize separately in your browser
See today’s decision
“VO₂max intervals, 60 minutes—should I still do it today?”
Planned session → today’s sessionWhat changed + whyConfidence + limits
Complete flow shown
Evidence folder structure
The installer asks for an optional private Evidence folder. Select the parent folder containing any supported export folders:
Your exported health data folder/ ├── export_apple_health/ ├── export_garmin/ ├── export_strava/ └── export_google_health/
Optional Google Health authorization
Ask “Connect Google Health” after installation. Open the returned link in your own browser, approve read-only access, then ask “Sync Google Health.” Pacevera normalizes the response in memory, and writes only normalized Evidence to the selected private folder.
Version and fingerprint verification
Download pacevera.mcpb from the release page. Check shasum -a 256 pacevera.mcpb against fileSha256 in the public MCP registry, then install it from Claude Desktop → Settings → Extensions.
Actual operation video · separate from the engine replay
A useful answer
is not yet a decision.
The same training question can produce a reasonable weather-based answer or a decision tied to the session already on your calendar. No Pacevera can find the temperature and offer useful advice. Pacevera adds the personal evidence, state, rules, and planned session → today’s session change.
No Pacevera
Finds external conditions such as temperature and turns them into a useful general training suggestion.
Pacevera Decision Engine
Reads the evidence supplied for this person, evaluates today’s state, and returns a traceable action from the planned session to the session to execute.
User scenario: “Threshold repeats, 60 minutes at high intensity today — should I run it?” The AI host understands the question. No Pacevera can check the weather. Pacevera checks the evidence and the scheduled workout, then answers whether to keep, adjust, substitute, defer, or advance today’s plan — with the reason, missing signals, confidence, and rule trace.
How it works · one decision, end to end
One person, one day,
one decision.
The brief above is the last step of a chain. Here is the whole of it — what arrived, what it computed, what it decided, what changed in the plan, and why. These are the values this page loaded, not an illustration: they come from the same engine output the brief is rendered from, so if the engine changes its mind, this section changes with it.
01EvidenceWhat the caller supplied. No account, no fetch.
sourcesapple_health, strava
window2026-07-12 → 2026-08-04
recovery signals2 usable · hrv, stress missing
training signals1 usable · none missing
02Fitness stateComputed from the signals that were fresh, weights renormalised across them.
readiness73
recovery79
fatigue27
03DecisionThe intent. One of keep, adjust, substitute, defer, advance.
typeadjust
intentreduce_today_intensity
04ActionThe same intent lands differently on different equipment, time or injury.
Planned session
VO₂max Intervals
→
Today’s session
Moderate run
changedfocus, durationMinutes, intensity
05ReasonTied back to a value and a versioned rule, so it still reads in October.
- 01Acute:chronic workload ratio 1.58 is above 1.4: load has been ramping too fast.
- 02At moderate intensity the session is no longer "VO₂max Intervals"; it becomes "Moderate run".
- 03Only 45 minutes are available, so the session is shortened.
ruleEVD-R-006 · An acute load spike pulls intensity down one step
measuredacute_chronic_workload_ratio = 1.58
thresholdacwrHigh > 1.4
confidencemedium
engine · rules1.12.0 · 1.9.1
Confidence is medium, not high, and the reason is on the first row: no training load arrived, so the load side of the picture is missing and the engine says so rather than filling it in. That is the whole difference between a decision you can act on and a sentence that sounds right.
Deep demonstration · continuity
What happened today
can inform what happens next.
A decision does not end when the answer appears. You can adopt it, change it, or skip it. That outcome can become local context for a later decision—but it does not train a model or rewrite Pacevera’s rules.
DC-001 produced an adjusted session. The user may adopt, change or skip it. Only the user’s actual response can become a local Outcome record. A future decision can read that record with fresh Evidence while the versioned rules remain unchanged.
DC-001 · Adjust
A reviewed synthetic engine replay supplies the decision. No outcome is invented for the case.
Adopt, change, or skip
Explore what each response would record. This page keeps the selection only for this demonstration.
LOCAL OUTCOME RECORDWaiting for your choice—Pacevera does not guess it.
Read context, then decide again
Fresh Evidence, the current plan and the locally held Outcome can be evaluated together. A full next-day verdict requires its own reviewed replay case.
Rules stay versioned.
A user response never teaches the engine to silently change a threshold or verdict.
A user response never teaches the engine to silently change a threshold or verdict.
Complete flow shown · no outcome selected
Your data boundary
Your health history
stays under your control.
The current Pacevera extension runs on your computer. It reads the folder you choose, computes locally, and returns the minimum decision record to your AI host. Pacevera has no account containing your health history.
Evidence enters locally
Describe a session, select a private export folder, or start the optional Google Health connection yourself.
Raw provider responses stay out
Google Health responses are normalized before disk. Tokens and health values do not enter MCP output.
A bounded record returns
The host receives the action, reasons, confidence, missing signals, and rule trace needed to explain today’s decision.
Your phone or desktop AIYou ask about the session already on your calendar. On mobile, the AI host’s remote feature reaches your own computer.
Pacevera · your computerRead allowed Evidence → standardize → calculate → apply deterministic rules
Your AI explainsThe bounded Decision + Action + Reason returns to the conversation.
If you use an AI platform’s mobile or remote feature, that platform remains part of its own data path. Your Evidence does not pass through Pacevera infrastructure.
AI Coaching
Your AI can coach. Pacevera helps it decide.
Use the AI host you already trust to ask questions and communicate in your own words. Pacevera takes the structured part: calculating fitness state, applying rules, and returning an action that can be checked later.
01
You ask your coach
“Should I keep today’s Threshold Intervals?”
02
Pacevera evaluates evidence
Recovery, fatigue, load, constraints and the scheduled workout.
03
You receive a decision
A concrete change to the planned session, with reasons and limits.
→
Planned session: Threshold Intervals
60 min · high intensity
→
Today’s session: Moderate run
60 min · moderate intensity
✓
Decision: adjust
Readiness 52 · high confidence
Decision Layer
Not another recommendation. A change you can act on.
Pacevera starts with the session you planned. It returns an accountable change — keep, adjust, substitute, defer, or advance — instead of inventing a workout from scratch.
Advice is cheap.
A decision has a prior state.
A general model knows the sports science. What it does not know is what you were supposed to do today, what your baseline HRV actually is, or that your sleep feed has been empty for a year. That gap is not a prompt problem.
A general AI coach
“You trained hard yesterday — take an easy Zone 2 run today.”
Where it comes fromInferred from language
Prior state neededNone
Auditable laterNo record of what drove it
A frontier modelCan do this today
Pacevera
“Your scheduled threshold intervals drop to moderate intensity — readiness 48, rule EVD-R-002.”
Where it comes fromYour evidence × a versioned rule
Prior state neededThe session you had scheduled
Auditable laterValues, rule id, missing signals, versions
A frontier modelCannot — it does not know today’s plan
Governance
It does not claim
to be right.
It shows you what it stands on.
Decisions rest on numbers somebody chose. Pacevera stores each threshold together with where it came from and returns that with the decision — and the rule library fails to load if a real paper is attached to a score we invented. 12 of 15 rules carry no source for their own threshold, and they say so.
Explainable
Every decision returns the rule that governed it, the threshold that rule cut, the value measured at that moment, and what each reason traces back to.
15 rules · 25 thresholds
Reproducible
The same evidence always produces the same decision, and every answer carries the engine version, the rule library version and its checksum — so you can tell which build said it.
same evidence in · same decision out
Auditable
No threshold moves until someone has seen what it did to the decisions. Every scenario is pinned to the answer that was approved, not to whatever the engine returned last time.
9 checks × 53 scenarios = 477 verdictsNo rule claims that a paper chose its threshold: 12 of 15 cite nothing at all, and the 3 that do cite literature record it as directional only.
EVD-R-006Externally defined metric
ThresholdacwrHigh > 1.4
Basisexternal_metric
Threshold sourceGabbett 2016
Verified toprimary full text verified
Also citedGriffin · abstract verified full text not read
Study design · rule gradesystematic_review
Supports thresholdsupports direction only
Loaded against itImpellizzeri 2020 · Lolli 2019
The paper says 0.8–1.3 is the sweet spot and ≥1.5 is the danger zone. 1.4 is neither — we chose it, and that sentence lives in the rule’s own limitation field. Citing Gabbett without loading his critics is the version that gets caught.
EVD-R-002Score we composed ourselves
ThresholdreadinessReduce < 60
Basisinternal_composite
Sources for the thresholdempty
Also citedDüking · abstract verified full text not read
Study design · rule gradenone
Supports thresholdinternal heuristic
Rules like this12 of 15
Readiness is our own weighted composite. No study has ever used this score, so its threshold can never have a citation. The empty sources field is an honest state, not an unfinished one.
Attaching a real study to an invented score is the worst mistake this product could make. So it is a load failure, not a guideline.
Each citation also carries how far it was verified — down to unverified. Making that field mandatory immediately surfaced a citation nobody had ever checked, and a review has since withdrawn claims rather than reword them.
The review log
A mechanism is worth what it catches on its own side.
Provenance you cannot see fail is a marketing claim. These are the times ours refused something we had already published.
2026-08-07
Three rule sources re-reviewed: one up, two down
The Gabbett citation matched an open full text and was upgraded. A claim that VO₂max drops 4–7% in two to three weeks appeared in neither abstract and sat behind a paywall — the passage was removed rather than rephrased.
2026-08-07
Verification level made mandatory
The same day, an unverified citation surfaced that had survived that review. While the field was optional, “never checked” and “nothing to note” looked identical in the output.
2026-08-08
The counter-evidence exemption was closed
Counter-evidence had been exempt on the grounds that we do not assert it — but it is still a claim about what a paper says, and readers do not draw that line. Both counter-citations on EVD-R-006 failed immediately; one turned out to be an editorial with no abstract at all. Both were downgraded, with the withdrawn text kept in the record.
Standing
A study on an invented score is a load failure
12 of 15 rules carry no source for their threshold. The library will not start if one of them cites literature.
Use cases
Six ways to ask.
Four kinds of answer.
Paste any of these after installing. Every output below came out of the engine and is pinned by a test — if this page and the code disagree, the test fails.
01 · No files, spoken evidence only
“I ran 80 minutes yesterday and felt wrecked, slept six hours. Today's plan is VO₂max intervals, 60 minutes. Should I still do it?”
Nothing was imported. One workout with no load figure and one recovery reading out of four — and the session still runs, with the confidence saying how thin the ground under it is.
02 · Strava only — zero recovery signals
“Here are my last four weeks of Strava activities. Threshold repeats, 60 min high intensity today — should I run it?”
Strava measures no HRV and no sleep. What it has is per-session load, and that alone moves the session. Ask where 1.4 came from and the trace answers with the citation, the published objections, and the admission that 1.4 is neither published number.
03 · Garmin only — vendor score as first-class evidence
“Here's my Garmin data for today. Tempo Run, 50 minutes, high intensity.”
The same decision type as the Strava case and the same one-step drop, for a completely different reason: load climbing too fast there, poor recovery today here. Body Battery is Garmin's own composite and is taken as evidence rather than recomputed — the watch was on the wrist and integrates signals we never see.
04 · Oura only — no load curve at all
“Here's today's Oura data. Low-intensity intervals planned for 60 minutes and I feel great — can I push?”
Recovery reads excellent and the answer is still keep. Oura computes no training load, so neither do we — multiplying duration by an intensity label would be easy, and that invented curve becomes a “back off today” three weeks later. The one rule that raises intensity withholds while confidence is low, so a strong recovery score on its own does not add load.
05 · Injury substitution — a hard filter
“My knee is bothering me. What can I swap today's squats for? I have a barbell, a rack and dumbbells.”
The returned reason names what was removed: movements contraindicated for the knee were hard-filtered out. Equipment was complete on purpose, so the knee is the only remaining reason for the swap. A model can be argued past a safety rule; a filter cannot.
06 · Whoop only — the session is removed
“Here's today's Whoop data. VO₂max intervals, 60 min high intensity — do I still do it?”
Five fields change, movements included. Adjust softens a session; defer replaces it. Confidence is high even though Whoop reports no stress — that cell stays empty, three recovery signals plus the vendor score are present, and that is enough to cancel a hard session outright.
Two of them attach no file at all. Importing data is not a prerequisite — and when you ask what to train with no plan on the table, the engine says that is a request for advice, not a decision, instead of inventing one.
Nobody has to complete the set.
Strava only? That is load. Oura only? That is recovery. Every combination produces a decision; the difference is which evidence it rests on and how much confidence it claims. Weights renormalise across the signals that are actually fresh — missing is never filled in with a neutral value.
| What you have | What it decides on | Decision quality |
|---|---|---|
| Strava only | Session load · ATL / CTL / TSB | Adjusts intensity and volume off the load curve |
| Apple Watch only | HRV · resting HR · activity | Adjusts off recovery signals and muscle-group fatigue |
| Garmin only | Body Battery · recovery time · load | Adjusts off the vendor composite and load |
| Oura only | Sleep · HRV · resting HR · readiness | Adjusts off recovery state, no load curve |
| Whoop only | Recovery signals + per-session strain | Both sides covered |
| Garmin + Apple Health | All of the above + HRV | The same decisions, at higher confidence |
Confidence is not accuracy. It describes how much fresh Evidence supports this decision and which signals are missing; it is not a success rate or a guarantee.
Adding a source never moves a decision from impossible to possible — only from lower to higher confidence. Which is why this product is not measured in connector count.
“The signal is there” never means that watch measured it.
Apple Health and Google Health are destinations as well as sources — other apps write into them. So “this came from Apple Health” tells you where it was exported from, not who measured it. Every reading therefore carries its writer and its last date, all the way through to the tool output.
Garmin syncs into Apple Healthsleep · resting HR · steps · workouts
The sync does not includeHRV
So the HRV in that exportall of it from an Apple Watch retired two years ago
hrv_msApple Watch · last 2023-08-03
sleep_duration_hoursGarmin Connect · last 2026-07-15
resting_hr_bpmGarmin Connect + Apple Watch
Ask only “is there HRV?” and the answer is yes — then a plan gets built on a signal this person stopped producing two years ago. It also lets missing have a type: a Garmin HRV field returning the same sentinel for 330 days is not a watch that cannot measure it, it is a watch that was not worn overnight. “Your device can’t” is a dead end; “wear it to sleep and it will” is actionable.
Who it is for
People who already have the data and will not hand it over again.
Not the general fitness market. Athletes and coaches already using an AI, already holding an Apple Health, Garmin or comparable export, whose open question is not “another dashboard” but whether today’s scheduled session should still be run.
First success is one traceable change to a plan, inside ten minutes.
The four questions everyone asks first.
Does it replace the AI I already use?
No. You keep asking your own AI in your own words; Pacevera takes the structured part — computing fitness state, applying rules, and returning a change that can be checked later.
Do I have to upload my health history?
No account and no upload. The engine runs on your machine, and evidence arrives as an export file, a training log, or simply what you type. Pacevera stores no Apple or Garmin credentials.
Where do the thresholds come from?
Each one is stored with its provenance and returned with the decision. 12 of 15 rules carry no source for their threshold. They are labelled that way — the library refuses to load if a real paper is attached to one of them.
What if I only have one data source?
Every combination still produces a decision. Weights renormalise across the signals that are fresh, missing signals are named, and confidence is reported honestly rather than inflated.
Public preview · $0
Start with the decision.
One path in, and it works right now on your own machine. Pacevera Desktop is the current install path while we validate the decision engine with serious athletes and coaches. No account, no health-data signup, and no email to hand over.
Pacevera Desktop · preview
Install now
$0
Available today
Pacevera v0.5.8
Pacevera Desktop is the current install path while we validate the decision engine with serious athletes and coaches. Everything runs from one file on your own computer — no Pacevera account and no API key. During installation, choose the private folder where local evidence and an optional Google Health grant will live.
- Local Pacevera MCPB for Claude Desktop
- Today’s session decisions: keep, adjust, substitute, defer, advance
- Evidence coverage, missing signals and confidence
- Decision reason and rule trace
- No account and no health-data upload to Pacevera
Before you click · two requirements
pacevera.mcpb · ~340 KB · macOS & Windows
The prerequisites are the whole gate — there is no account to create
Pacevera
Every training decision deserves a reason.
Your readiness score is only the beginning. Pacevera shows what changes in today’s planned session, why it changed, and what evidence is still missing.
Three months from now, you should still be able to ask why today changed: which rule fired, which values triggered it, what was missing, and which version of the engine and rule library produced it.
Pacevera provides general fitness decision support. It is not medical advice, diagnosis, treatment, emergency guidance, medical clearance, or a medical device. Read the Privacy Policy and Terms of Use.
Get started for free