getbased — personal health intelligence, under your control

getbased is an open-source health dashboard for people who want to understand their own biology without handing the whole record to a black-box health app. It brings labs, DNA, wearables, light exposure, lifestyle context, notes, and optional AI analysis into one browser-based workspace.

You can use it with no account. Most data lives in your browser by default. Health-data network features — AI providers, encrypted sync, profile sharing, external Knowledge Bases, and Agent Access — are opt-in. On the hosted app, cookieless anonymous pageview and outbound affiliate-click counts are enabled by default; they contain no health data and can be disabled in Settings → Privacy.

Live app · Documentation · Discord · Nostr

getbased Dashboard

• Import lab reports — drop PDFs, images, spreadsheets, or manually enter values. getbased maps known markers, lets you review values, laboratory ranges, collection time, and fasting status before saving, and keeps reviewable snapshots per file.
• Track biomarkers over time — 196 core markers across 19 categories, plus specialty and custom markers, with charts, tables, heatmaps, lab-first and context-aware ranges, optional evidence-backed optimal ranges, notes, trend flags, and date comparison.
• Use calculated markers without duplicates — supported lab-reported ratios and indices take priority, while deterministic fallbacks cover lipid/metabolic ratios, NLR/PLR/MLR/SII, FIB-4, anion gap, free water deficit, and biological age when their inputs are available.
• Read biology as patterns, not isolated numbers — Biology Scores summarize deterministic marker patterns such as metabolism, thyroid, cardiovascular health, inflammation, methylation, iron/blood, hormones, stress resilience, cellular energy, gut-immune terrain, and Biological Coherence. They are pattern summaries, not diagnoses.
• Bring in DNA context — raw DNA imports from common consumer and clinical formats, with curated SNP interpretation, APOE haplotype support, mtDNA haplogroups, and DNA-aware AI context.
• Connect wearables and body metrics — the official app supports Oura, Withings, Polar, existing legacy Fitbit connections, local Apple Health file import, and manual weight, blood pressure, and resting pulse. The hosted cloud integrations use narrowly scoped provider relays; WHOOP, Ultrahuman, and Google Health remain self-host only and use infrastructure controlled by that deployment. Every stored OAuth token is device-key encrypted.
• Track light and environment — sun sessions, UV/atmospheric context, indoor light setup, devices, measurements, EMF assessment, and daily light analysis.
• Keep a complete therapy history — track current, scheduled, paused, cycled, and ended supplements or medications with dose periods, structured units, active and other ingredients, reviewed product-label imports, and source quality evidence.
• Add the missing human context — medical history, family history, diet/digestion, sleep, exercise, stress, light/circadian habits, environment, EMF, health goals, and freeform notes. Cycle tracking can label individual blood draws and apply phase-specific estradiol, progesterone, LH, and FSH ranges only when context is reliable.
• Ask AI with context — chat can use your labs, notes, bounded therapy history, scores, wearables, Knowledge Base passages, and selected interpretive lens. Dictate into the composer and listen to replies with browser-local voice models, a local compatible server, a supported OpenRouter/PPQ/Venice AI connection, xAI, or ElevenLabs.
• Build reports — export a practitioner-readable PDF with selected labs, context, and summary sections.
• Use multiple profiles — separate profiles for yourself, family, clients, or test/demo data.
• Protect and move your data — create full backups, optionally encrypt browser storage with a passphrase, sync encrypted profiles across devices, or share a password-protected copy.

getbased is private by default, not magic. The boundary depends on which features you turn on.

• No account required. You can open the app and work locally.
• Browser-first storage. Profile data is stored in localStorage and IndexedDB by default.
• Optional encryption at rest. A passphrase-derived key can protect browser storage.
• Optional AI. PDF import and chat need either an AI provider or a local OpenAI-compatible server. Non-AI tracking features still work without one.
• One activation decision, separate records. The first remote AI activation presents the provider-neutral AI notice and destination-specific approval together, while storing them separately. It explains that enabled automatic insights may request updated analysis after relevant profile or data changes. Same-device inference needs only the AI notice. A private/LAN endpoint is confirmed by origin, while each remote provider or endpoint needs a browser-local sensitive-data approval before any data-bearing AI request can proceed. User-triggered connection checks may run first, but contain no profile, chat, image, or voice content; a custom endpoint may receive a fixed synthetic compatibility probe.
• Optional Voice. On-device Whisper/Kokoro keep recordings and message text in the browser after model download. A selected local server receives them directly. A selected OpenRouter, PPQ, Venice, xAI, or ElevenLabs audio endpoint receives only the recording or reply text explicitly processed with that cloud provider; Automatic can reuse a connected voice-capable direct AI provider but stays on-device during CLI chat unless another voice service is explicitly selected.
• PII review for text imports. Deterministic patterns and an optional trusted self-hosted model can strip likely identifiers before lab text is sent to an AI provider. Automated detection can miss unusual layouts, so review is still recommended. Image imports cannot be scrubbed and always show a separate warning before upload.
• Optional encrypted sync. Cross-device sync uses Evolu CRDT storage and end-to-end encrypted profile payloads. Pausing one browser keeps its identity and queues local edits; disconnect/reset is a separate Advanced action.
• Optional sharing. Profile sharing creates an encrypted, password-protected copy for someone else. On the official app, getbased's isolated Czech VPS service (hosted by SecurityNet.cz/Hukot) receives caller network metadata plus the opaque envelope and expiry/deletion/abuse metadata, but not the password or decrypted profile. Temporary share copies are not backed up and can be lost after a service failure; the source profile remains in the sender's browser.
• Optional Agent Access. External agents receive only the context you enable, via an encrypted relay flow and a local decryption key.
• Optional wearable clouds. Every connection is user-initiated. On the official app, fixed Oura, Withings, Polar, and legacy Fitbit OAuth/API calls transit a separate getbased compatibility relay without intentional payload logging or storage; the relay can nevertheless read those credentials and provider responses while forwarding them. The static Vercel app host does not receive these compatibility requests. WHOOP, Ultrahuman, and Google Health are code-enforced self-host-only integrations and use infrastructure controlled by that deployment. Confidential token exchange and refresh use its same-origin proxy; WHOOP and Google Health data also transit that proxy, while Ultrahuman data is fetched browser-direct because its resource API permits browser CORS. OAuth credentials are encrypted with a non-exportable device key; WHOOP and Google Health daily rows use the same device-only protection and are excluded from raw-data backups. WHOOP-specific connection metadata and derived local values are likewise split into device-encrypted storage. Each browser must be connected separately; only the compact derived summary participates in optional end-to-end-encrypted sync or user-enabled AI/agent context.
• Minimised Sun/UV relay. On the official app, CAMS requests are forced to a 0.1° grid in the browser and again in the separate compatibility relay, then sent by authenticated POST to a fixed Company-operated local-grid lookup. The route does not forward or cache individual coordinates at Copernicus or Open-Meteo. Missing weather fields fall back browser-direct to Open-Meteo with the same rounded location. This is limited plaintext processing, not end-to-end encryption.
• Hosted-app usage stats. Cookieless pageview and outbound affiliate-click counts are enabled by default with a first-run transparency banner and can be disabled immediately or later in Settings → Privacy. Health data, viewed records, identity, and health context are not analytics payloads.

If you want the strictest setup, use an AI server on the same device, disable analytics, and leave sync, sharing, Agent Access, connected wearable integrations, external Knowledge Bases, and other remote data sources off.

All normal tracking works without AI. AI features can use:

Switch providers in Settings. Supported provider keys are wrapped locally with device-bound encryption, or with the backup passphrase when one is configured. The activation prompt identifies the selected recipient and links its Privacy and Terms documents when known; reviewing those links is not an acceptance of the provider's terms on its behalf. Custom endpoints are identified by their origin. Users connecting a personal or local endpoint are not asked to supply policy metadata.

On Linux, macOS, and Windows, Settings provides one OS-specific bootstrap command that runs the companion temporarily. Once connected, Settings can directly register automatic startup, pause or resume AI work, restart Codex, update the installed bundle, or uninstall it. A recovery start command remains available when the local process is completely stopped. The bootstrap downloads the same auditable, single-file getbased-companion.mjs emitted by the production build. Temporary mode stops with its Terminal or PowerShell session. Installed mode uses a systemd user service, macOS LaunchAgent, or current-user Windows scheduled task. It does not need root or administrator access, and the browser never exposes a port or pairing token. The entry point, host, installers, protocol, and bundle build are all source-available in this repository under bin/, server/, lib/, shared/, and scripts/.

The dormant Claude adapter is intentionally not advertised or detected in normal builds. Anthropic does not permit third-party products to route claude.ai Free, Pro, or Max credentials without prior approval. Self-hosting operators using Anthropic Console/API billing may expose the adapter as Claude Agent by starting or installing the companion with GETBASED_ENABLE_CLAUDE_AGENT=api-console; then sign in with claude auth login --console. This opt-in does not make consumer subscription credentials permissible.

Independent deployment operators can identify themselves without inheriting getbased policies by setting these metadata values in index.html:

<meta name="getbased-operator-name" content="Example Health Cooperative">
<meta name="getbased-operator-privacy-url" content="https://example.org/privacy">
<meta name="getbased-operator-terms-url" content="https://example.org/terms">

Empty values do not fall back to getbased Privacy or Terms. The official getbased host supplies its own links automatically, as supplementary links in AI destination prompts; the selected AI recipient's documents remain primary. An operator that needs provider metadata beyond the built-in catalog can also define GETBASED_DEPLOYMENT_CONFIG.aiProviders[providerId] with label, privacyUrl, and termsUrl before the main module loads.

Venice encrypted mode fails closed unless both the Intel TDX quote and NVIDIA GPU evidence verify. NVIDIA does not allow browser POSTs to NRAS, so the fixed GPU-attestation request uses the compatibility endpoint selected for the deployment. Official hosts route it to integrations.getbased.health; independent deployments use their own same-origin /api/proxy. The policy permits only NVIDIA's exact NRAS endpoint for this operation and does not receive the encrypted inference messages. Signed NRAS tokens are then verified in the browser against NVIDIA's published keys. Message content is not included in that evidence. The shared nonce does not prove that the GPU and TDX workload are co-located.

Voice input and output use one service by default in Settings → Voice. An advanced toggle can select different services for dictation and spoken replies, for example on-device Whisper input with ElevenLabs output:

• Automatic reuses a connected OpenRouter, PPQ, or Venice account during direct-provider chat when that provider supports the requested audio operation, and otherwise falls back to the browser-local models. During CLI chat it always stays on-device unless the user explicitly selects another voice service. A selected CLI agent receives only the transcribed text, not the recording.
• On this device uses quantized multilingual Whisper Small by default, with Medium as a balanced accuracy step-up and Large v3 Turbo as the high-end option. Local transcription and Kokoro speech can use CPU/WASM or a validated WebGPU path. Automatic processing tries WebGPU first on capable mobile browsers, starts with the broadly compatible CPU path elsewhere, falls back safely, records normalized timings, and then selects the fastest tested backend. Each required CPU or GPU model variant downloads only after explicit confirmation and runs in a dedicated Web Worker. Slow transcription shows elapsed time and can be cancelled; a stalled local transcription is stopped after three minutes. Kokoro buffers short opening segments to avoid a long gap while it generates the next sentence. Built-in speech currently offers English US and UK voices. Markdown tables are skipped with a short spoken notice so the surrounding explanation remains easy to follow.
• Local voice server connects directly to an OpenAI-compatible /v1/audio/transcriptions and /v1/audio/speech server, including apps such as LocalAI or Speaches when configured with compatible models.
• OpenRouter, PPQ, and Venice reuse their key from AI settings and call the provider's standard audio endpoints directly. A private chat mode does not automatically extend to voice.
• xAI and ElevenLabs use your own API key. There is no delegated account sign-in for these integrations; supported provider keys are device-key encrypted even when passphrase protection is off.

Microphone audio is held only for the active transcription. Dictation inserts text at the composer cursor and never sends a message automatically. Assistant messages expose a Listen control, and recording or playback stops when the chat closes or switches conversations.

The Knowledge Base lets AI ground answers in your own documents instead of relying only on model memory. It supports:

• an in-browser local library for documents indexed on this device;
• an external knowledge server for larger or shared libraries;
• citations/snippets injected into chat and Current Focus when relevant.

The Interpretive Lens is different: it changes the framing of analysis. For example, you can ask the AI to read the same labs through a mitochondrial, endocrinology, circadian, or other scientific lens.

Agent Access is for using your getbased context from external AI tools — Hermes Agent, OpenClaw, Claude Agent, Claude Desktop, Cursor, Cline, Codex CLI, or another MCP-compatible client.

How it works:

1. Enable Cross-device Sync.
2. Enable Settings → Agent Access.
3. Pick the target client.
4. Copy the private setup command.
5. Paste it into that agent's terminal.

The public installer provides one-command setup on Linux:

curl -fsSL https://getbased.health/install.sh | bash

This installs the local agent stack only. On macOS, Windows, and WSL1, follow the manual Agent Access setup instead; the installer cannot configure systemd user services there.

Private access requires the setup command copied from the app:

curl -fsSL https://getbased.health/install.sh | bash -s -- connect <target> --setup 'gbsetup_v1_...'

That setup payload contains the read-only relay token and the local Agent Context key. Do not paste real setup values into logs, issues, or public docs.

Requires Node.js 24.

git clone https://github.com/elkimek/get-based
cd get-based
npm ci
npm run dev-server

Open http://localhost:8000/app. The root URL may serve the sibling get-based-site landing page when that repository is present.

Useful checks:

npm run typecheck
npm run typecheck:checkjs
npm run architecture:check
npm run vendor:check
npm run supply-chain:check
npm run sbom
npm run quality
npm test -- tests/<relevant-test>.test.js
npx playwright test tests/playwright/<relevant-spec>.spec.js
npm run test:evolu8-browsers
npm run test:firefox
npm run performance:check
npm run production:check

Default to tests related to the current change. GitHub Actions runs the exhaustive browser and combined-coverage matrix so local development does not repeatedly create high volumes of temporary Chromium and V8 coverage data. ./run-tests.sh runs both type checkers, verifies the architecture map, vendored browser assets and their supply-chain inventory, and the static module graph, starts an isolated local server, runs the Node/Vitest tests, checks the dev-server origin guard, and runs every Playwright browser assertion. It is blocked outside CI unless the high-write run is explicitly acknowledged with GETBASED_ALLOW_HIGH_WRITE_TESTS=1. COVERAGE=1 ./run-tests.sh also combines Vitest and Playwright V8 function coverage and enforces the committed ratchet in scripts/coverage-baseline.json; CI runs this mode on every change. npm run test:evolu8-browsers runs the focused Evolu 8 startup, durable-identity, resource-management-polyfill, and one-tab fallback checks in Chromium, Firefox, and WebKit. npm run test:firefox runs the focused Firefox critical-flow suite; install its browser binary once with npx playwright install firefox. npm run performance:check runs the focused cold mobile-load check and enforces the committed request-count, compressed-transfer, and decoded-byte ceilings. npm run production:check builds the deploy artifact in a temporary directory and enforces the production startup, lazy-chunk, and PWA app-shell precache resource/decoded-byte budgets without changing the worktree. npm run sbom writes a combined CycloneDX inventory for npm and vendored browser components to artifacts/getbased.cdx.json.

• Native browser ES modules in source; the Vercel build uses Rolldown to collapse the static startup graph while preserving feature-level lazy chunks.
• Plain HTML/CSS/JS with split modules under js/ and feature CSS under css/.
• Chart.js for charts.
• pdf.js for PDF text extraction.
• transformers.js + OPFS for the in-browser Knowledge Base.
• transformers.js, quantized Whisper Small/Medium/Large v3 Turbo, and Kokoro for optional in-browser voice.
• Evolu for optional encrypted CRDT sync.
• A getbased-operated, SQLite-backed service for opaque encrypted profile shares; a Vercel endpoint for public deployment metadata; and a separately deployed, narrowly scoped compatibility relay for supported wearable providers, NVIDIA attestation, the fixed privacy-rounded CAMS lookup, and credential-free public-page imports. Generic AI, voice, and custom-provider forwarding is rejected on getbased-operated hosts. Self-host-only providers use the deployment owner's OAuth credentials and, only where required, its same-origin proxy.
• Vitest, TypeScript checkers, quality guardrails, and Playwright for verification.
• PWA install support; core local tracking and cached app data work offline. Integrations, synchronization, sharing, remote Knowledge Bases, and live environmental data require connectivity.

get-based/
├── index.html, styles.css, css/ # App shell and feature styles
├── service-worker.js # PWA caching, offline shell, and network routing
├── js/ # Native ES modules
│ ├── settings-agent-access-panel.js
│ ├── sync*.js # Optional encrypted sync and Agent Access plumbing
│ ├── lens*.js # Knowledge Base backends and query injection
│ ├── biology-score*.js # Biology Scores engine, UI, and AI context
│ ├── wearables*.js # Wearable adapters, storage, settings, summaries
│ └── light*.js / sun*.js # Light & Sun tools, sessions, environment, AI summaries
├── api/ # Vercel/serverless routes
├── lib/ # Node-only server policy and transport
├── data/ # Curated marker, SNP, lens, and reference data
├── brands/ # Product and provider brand assets
├── scripts/ # Architecture, vendor, catalog, and maintenance tooling
├── tests/ # Vitest and Playwright coverage
├── vendor/ # Vendored browser libraries
├── ARCHITECTURE.md # Maintained ownership and dependency contract
├── MODULE_MAP.md # Generated runtime module/import map
└── .github/workflows/ # CI

User and developer documentation live at docs.getbased.health. The app repo keeps only code-adjacent notes and tests. The renderer-independent report snapshot contract is documented in docs/report-data.md.

• getbased-docs — public user and developer documentation.
• getbased-agents — MCP adapter, local knowledge server, dashboard, and getbased-stack installer.
• getbased-relay — Evolu sync relay and Agent Access context gateway.
• get-based-site — landing page, public installer, llms.txt, and agent discovery files.

See CONTRIBUTING.md. Project board: planned features.

AGPL-3.0-or-later. See LICENSE.

If you run a modified version as a network service, AGPLv3 §13 requires you to offer users the corresponding source. Vendored third-party libraries are listed under their own licenses in THIRD_PARTY_LICENSES.md.