Pacevera

Pacevera Fitness Decision Engine

Pacevera is a deterministic fitness decision engine exposed through the Model Context Protocol (MCP). It receives normalized training and health evidence supplied by the calling AI host, evaluates recovery and training constraints, and returns a structured decision.

The primary operation is to evaluate an existing scheduled session. A decision describes the change from the scheduled session to the resulting session:

scheduled session (from) -> resulting session (to)

Each decision includes the relevant evidence, applied rules, confidence, signal coverage, and limitations.

Scope

Pacevera provides the following capabilities:

• Deterministic calculation of recovery state, readiness, muscle-group fatigue, and training load.
• Session decisions with the types keep, adjust, substitute, defer, and advance.
• Constraint handling for injuries, equipment, available time, and exercise substitutions.
• Training-plan generation and caller-owned plan adjustment through preview and commit operations.
• Structured decision provenance through rule identifiers, measured values, source information, and version metadata.
• Normalization support for Apple Health, Google Health Takeout, Garmin, Strava, Oura, and WHOOP data shapes.

Pacevera does not connect to these providers on behalf of the user. The calling host must obtain and normalize provider exports or API responses before passing them to an MCP tool.

Architecture

AI host  └─ Collects the user's question and evidence       └─ MCP server / desktop extension            ├─ Evidence normalization            ├─ Fitness-state calculation            ├─ Training-load calculation            ├─ Deterministic rule evaluation            ├─ Decision and plan operations            └─ Structured result: Decision / Action / Reason

The AI host is responsible for interpreting the user's request, collecting evidence, selecting tools, and presenting the result. Pacevera performs the calculations and rule evaluation; it does not call a model to generate a decision. Given the same code, rule set, and input, the result is deterministic.

Decision and recommendation

decide_session requires scheduledSession to establish the prior state. If no scheduled session is provided, the tool returns no_scheduled_session rather than creating a session. Use generate_plan when a plan is required.

Decision output follows five layers:

Evidence -> Fitness State -> Decision -> Action -> Reason

Decision intent and resulting action are separate fields. The same intent may produce different actions depending on the user's equipment, available time, injuries, and scheduled session.

MCP tools

Decision and planning tools

Supporting and read tools

Evidence and output

Evidence is supplied by the calling host. It may include recovery measurements, vendor-computed assessments, completed workouts, scheduled sessions, goals, and constraints.

Important input rules:

• Only measured or explicitly provided values should be sent. Missing signals are reported in signalCoverage and may lower confidence.
• A workout without a training-load value is not treated as zero load and is excluded from muscle-group fatigue calculations.
• RPE may be retained as evidence but is not a term in training-load or muscle-fatigue calculations.
• Vendor-computed values such as readiness, recovery, and Body Battery are used as reported and are not recomputed by Pacevera.

A typical decide_session response has this shape:

{  "decision": { "type": "adjust", "intent": "reduce_today_intensity" },  "action": {    "from": { "focus": "Tempo Run", "durationMinutes": 50, "intensity": "high" },    "to": { "focus": "Moderate run", "durationMinutes": 50, "intensity": "moderate" },    "changed": ["focus", "intensity"]  },  "confidence": "high",  "signalCoverage": {    "recovery": { "usable": ["readiness"], "missing": ["sleep"] },    "training": { "usable": ["trainingLoad"], "missing": [] }  },  "decisionBasis": {    "governingRule": {      "ruleId": "EVD-R-002",      "measured": { "quantity": "readiness_score", "value": 48 }    }  }}

Pacevera does not fill missing values with defaults. signalCoverage.recovery describes current recovery-signal availability; signalCoverage.training describes whether recent workouts include the training-load data required for fatigue calculations.

Deployment status

The desktop extension has no runtime dependencies beyond the Node.js standard library. It does not use analytics, telemetry, crash reporting, provider accounts, or provider OAuth tokens.

Privacy

The desktop extension processes the minimum health-related Evidence supplied by the caller for the requested fitness decision. It does not retain, sell, use for model training, or use Evidence for unrelated purposes.

The extension makes no outbound network requests and sends Evidence nowhere. On the user's own computer it does keep durable records: a bounded continuity record, and a local SQLite store holding the decisions it produced, the outcomes the user reports back, and the derived state behind them. Both are the user's to export or delete. The AI host, operating system, imported files, and any host conversation history are outside the extension's control and are governed by their respective policies.

See PRIVACY.md for the complete policy and deployment scope, and TERMS.md for the terms of use. The canonical versions are published at https://pacevera.com/privacy and https://pacevera.com/terms.

Pacevera is not a medical device and does not provide medical advice. It is intended for general fitness and training purposes only.

Installation

Pacevera is distributed as a Claude Desktop extension (.mcpb).

1. Download pacevera.mcpb from Releases.

2. Optionally verify the SHA-256 checksum published in the release notes:

   bash

   shasum -a 256 pacevera.mcpb

3. In Claude Desktop, open Settings -> Extensions and install the file.

4. Restart Claude Desktop after installing or upgrading the extension.

The extension requires Node.js 20 or newer. Local persistence of decision traces and outcomes needs Node.js 22.5 or newer, which is where node:sqlite exists; on Node 20 or 21 the extension installs and runs without it, and reading your selected export folder for a decision is unaffected. The local plan decision is offered only once this machine holds a plan and its context, so a fresh installation does not list it — that depends on what is stored, not on the Node version.

Claude Desktop and other local MCP hosts are supported through the desktop extension, and Claude's own Dispatch can drive that installed extension from a phone with no remote deployment anywhere in the path — it needs the desktop app running, a Pro or Max plan, and the feature to be available to you. A Pacevera-operated hosted remote deployment is still not available for public use.

Release status

The current public release is v0.5.8, using Decision Engine 1.12.0 and [email protected]. It includes ten public tools, bounded local continuity, personalized single-workout generation, decision traces, package validation／dry-run／rollback, and an optional local Google Health connection. The Oura and WHOOP readers were implemented from their published API specifications and have not yet been validated against real responses; Apple Health, Garmin, Google Health, and Strava readers were developed from real export files.

Support

• Bug reports and technical questions: GitHub Issues
• Privacy requests: [email protected]

License

Pacevera is proprietary software. See LICENSE for the applicable terms.

Pacevera v0.5.8

The Claude Desktop extension includes Today’s Brief, evidence-first local export reading, and plain-language tool titles. During installation, choose Your exported health data folder and select the parent folder containing any of these optional subfolders:

Your exported health data folder/├── export_apple_health/        # Apple Health export.xml├── export_garmin/              # Garmin Export Your Data / DI_CONNECT├── export_strava/              # Strava activities.csv└── export_google_health/       # raw/ for JSON you pulled yourself,                                # normalized/ for what the connection wrote

Pacevera reads the selected folder locally; missing sources are reported as unavailable rather than guessed. The MCPB checksum is published in the release notes.

v0.5.8 includes an optional Google Health connection that runs on your computer. You approve it in your own browser, or by scanning a QR code with a phone; the browser route reaches Google with no Pacevera server anywhere in its path, and the QR route passes through a Pacevera relay that handles only a short-lived authorization handoff — never your health data, your tokens, or the PKCE secret that completes the exchange on your machine. Whichever route you finish cancels the other.

In v0.5.8 the connection normalizes each Google Health response in memory and writes only the resulting evidence; the provider response itself is never written to disk. You can also disconnect from inside the app: it revokes the grant at Google, deletes the stored credential, deletes the evidence earlier syncs wrote, and clears the derived records from the local store. Files you placed in the folder yourself are reported back for you to delete, never removed for you.

This connection is capped at 100 accounts, and Google will warn you that it has not verified the app. Pacevera's Google app moved out of Testing into Production on 2026-08-20, so an invitation is no longer required — but Google's verification review for the health scopes is not complete. Until it is, Google shows its own "Google hasn't verified this app" screen before you can consent, and at most 100 accounts can authorize the connection, counted over the app's entire lifetime. Reading your own exported folders, supplying evidence through your AI host, and every decision tool are unaffected.

v0.5.8 corrects three things a decision was saying that it could not support. When a movement you ruled out emptied a session, the replacement Pacevera offered had never been checked against the same restriction — say you cannot squat, and both barbell squats came out while a bodyweight squat went in, with the reason line mentioning only what had been removed. The replacement now passes the same check, and when nothing passes, no movement is named and the limitation says why instead of inventing one. It also tells you the movement is a fallback rather than something taken from your own plan, which it never used to mention at all. When no rule acts, the closing sentence no longer says every reading is in range: recovery below 60 on an already-easy day, or muscle-group fatigue at 40 on a day that is not hard, can cross a line without any rule acting, and the sentence now names the reading, the line, and why nothing stepped in. And on a week with no recorded training at all, it no longer says the target muscles carry no load from the last week — there was nothing to read, which is a different statement, and it is the one you now get.

Today's Brief follows the language of the conversation: ask in Traditional Chinese and the card is in Traditional Chinese, ask in English and it is in English. In v0.5.7 it followed your desktop app's display language instead, regardless of what you were speaking.

Pacevera also stops implying that Garmin has a connection to repair. Garmin readings only ever arrive as an export you place in your folder — there is no account to link and no sync to check — and the assistant is now told so, rather than offering to look into a sync that does not exist.

v0.5.7 fixes two ways the extension could stop answering, both of them present in v0.5.6. On a day when no fresh recovery reading arrived — the day a decision is most worth asking for — asking for today's decision from a stored plan returned nothing at all, because the local store refused to record the empty readiness the engine deliberately reports. Existing stores are repaired in place, with your rows intact. And a single very large, deeply nested request could end the extension outright, which looks from the outside like the whole connector vanishing; one bad request now fails on its own and the extension keeps running. The one rule that raises a session's intensity no longer raises it while confidence is low: thin evidence may hold a session or lower it, not add load. Coverage is also reported more plainly — the vendor scores that were actually used are named, and the coverage figure is a fraction of five kinds of evidence rather than a number that could run past its own maximum. The folder reading records what it imported from Garmin without handing your folder path or file fingerprints to your AI host. (This release was described as making Today's Brief follow the language you are speaking in. That was inaccurate: in v0.5.7 the card follows your desktop app's display language. v0.5.8 is the release where it follows the conversation.)

v0.5.6 reads your HRV against your own recent nights rather than a fixed reference, and where your device already computed a readiness or HRV score, that score is used instead of scoring the raw reading a second time. The raw reading is kept and reported as superseded, named alongside what replaced it, so a decision can no longer say a reading is missing when you supplied one. Connecting Google Health, syncing it, and asking for today's decision from a stored plan no longer ask you for an account identifier: Pacevera never issued one, and being asked for it again in a later conversation is what made a working connection look expired. Where this machine holds more than one profile, Pacevera asks which one by name rather than guessing.

v0.5.2 added the source chain to every reading — where Pacevera read it, and which app wrote it, so a Garmin figure synced into Apple Health is not reported as Apple Health's own. A reading whose writer the export does not name is reported as unnamed rather than guessed. Decisions also carry a small continuity record that a later conversation can hand back; Pacevera checks whether it still describes your current state and says so when it does not. It is checked, never merged: the decision always comes from the evidence held now.