# HugoScore CAIHL Draft Report: MyChart Takeout

- Status: Draft for human review
- Last reviewed: 2026-08-14
- Review method: Public-source review of the official landing page, the GitHub repository README, the package manifest, the repository layout, output-format documentation, and parity-test provenance notes, plus web context on the maintainer's public role. No hands-on export with a real MyChart account, security audit, real-PHI code execution, maintainer interview, accessibility review, or legal determination was performed.
- Service: MyChart Takeout
- Vendor: Josh Mandel (jmandel), independent open-source project
- URL: https://joshuamandel.com/mychart-takeout/ (repository: https://github.com/jmandel/mychart-takeout)
- Category: Patient-controlled health records AI

## 1. Executive Summary

MyChart Takeout is a free, source-public tool that exports a patient's complete Epic/MyChart record into a single local ZIP. The primary path is a bookmarklet: sign in to MyChart as usual, click the bookmark, click Start export, download the ZIP. A CDP-driven local CLI serves power users, adds forensic capture, and can export every proxy subject at once. The export replays MyChart's own internal JSON API calls from inside the patient's authenticated browser tab, so no server, account, or password handling is involved anywhere in the data path.

The exporter is explicitly deterministic. No language model touches the data path. Its HugoScore relevance is as AI-enabling patient agency infrastructure: it converts portal access into complete, local, structured, AI-ready files, then leaves every downstream choice, including whether and which AI to use, with the patient.

- Agency posture: Strongly agency-expanding
- Agency axis position: 92 of 100
- Confidence: Medium draft, source-public, no independent audit

## 2. CAIHL Question

**Who does MyChart Takeout serve?**

Patients, caregivers with proxy access, and patient-developers. The tool has no institutional customer, no account system, no telemetry, and no visible monetization. The landing page's framing is the project's alignment statement: "Your records, in your hands." It is independent of and expressly not affiliated with or endorsed by Epic Systems. The maintainer is a widely known health-interoperability architect, and the project reads as public-interest infrastructure rather than a product.

CAIHL classification: Patient-directed, AI-enabling record export utility.

## 3. What The Service Does

- **Bookmarklet mode (primary for patients):** a self-contained script runs on an open, signed-in MyChart tab, collects data with in-page fetch calls, and hands the patient a downloadable ZIP. No install, no backend.
- **CDP mode (power users and agents):** the patient launches Chromium with remote debugging, signs in, and a local CLI attaches over the Chrome DevTools Protocol, adding raw network logs, DOM snapshots, optional screenshots, and multi-proxy export.
- **Capture scope:** problems, allergies, medications, immunizations, histories, lab and imaging results with reference ranges and narratives, visit notes and After-Visit Summaries, every secure-message thread, standards C-CDA documents, insurance and coverage, implanted-device identifiers (UDI, model, serial), patient-tracked vitals, referrals, growth charts, connected apps, notification settings, and the access log showing who viewed the record.
- **Report layer:** each export opens with a derived PATIENT_SUMMARY.md, flat CSV indexes, and a MANIFEST.json, so the result is readable by a person before it is consumed by any program.
- **AI stance:** the README states the exporter is deterministic, with no language model in the export path. Agents can drive the CLI, and the structured output is deliberately AI-ready, but AI entry is always the patient's choice.

## 4. Patient-Impact Pathway

1. **Direct use:** a patient signs in to their own MyChart, runs the bookmarklet, and walks away with their complete record as local files.
2. **Caregiver use:** proxy support extends the export to children and dependents a caregiver manages, with each ZIP named for its patient.
3. **Evidence use:** exported notes, messages, and the access log give patients durable material for disputes, second opinions, advocacy, and record-correction requests.
4. **AI-enabling use:** the structured JSON, CSV indexes, and summary markdown feed whatever downstream AI tool the patient selects, on the patient's own machine and terms.

## 5. Evidence Table

| Source | Type | Claim supported | Notes |
| --- | --- | --- | --- |
| https://joshuamandel.com/mychart-takeout/ | Official landing page | Bookmarklet flow, capture list, in-browser-only architecture, "data never leaves your machine," no-password design, proxy support, Epic non-affiliation | Static page; bookmarklet is self-contained. Verified 2026-08-14. |
| https://github.com/jmandel/mychart-takeout (README) | Official repository | Deterministic no-LLM export path, internal-API replay design, CDP and in-browser modes, output layout, PHI written only to local git-ignored directories, known limitations | Verified 2026-08-14. |
| Repository package manifest | Official repository | Workspace layout; `"private": true`; no license field | No LICENSE file found at common paths despite the landing page's open-source claim. Verified 2026-08-14. |
| README parity provenance (P5, 2026-08-13) | Official repository | TypeScript exporter verified record-for-record against the prior Python implementation on a live session; 110-test suite runs against a synthetic-patient mock with no PHI | Self-verification by the maintainer, not an independent audit. |
| README known-limitations section | Official repository | Endpoint paths and features vary across Epic instances; FHIR R4 requires separate OAuth2 credentials; EHI export disabled on some instances | Basis for the instance-variability caveat. |

## 6. Mixed HugoScore Profile

### 6.1 Agency Posture (summary judgment)

**Strongly agency-expanding.** The tool moves a patient's complete Epic portal record into local files the patient fully controls, with no server in the data path and no language model in the export. The caveats are Epic-instance variability, an unstated software license, and the baseline requirement of portal access plus bookmarklet comfort.

### 6.2 Service Alignment

**Patient-directed and AI-enabling.** Primary user: patients, caregivers with proxy access, patient-developers. Hosting and control: a static public page plus a self-contained in-page script; the only runtime is the patient's own browser, and the only output destination is the patient's own disk. No vendor economics are visible anywhere in the project.

### 6.3 Patient Visibility. Can patients tell AI is involved? **Yes**

Disclosure is explicit and unusually precise: the exporter is deterministic and no language model is involved in the export itself. Agent-driven CLI use is documented as an option the user chooses, not a hidden layer.

### 6.4 Choice and Refusal. Can patients meaningfully choose? **Yes**

Use is entirely voluntary and patient-initiated, and doing nothing changes nothing. The bookmarklet path asks only that a patient drag a button to the bookmarks bar, sign in as usual, and click Start, a far lower activation burden than token- or script-based alternatives in this category. The deeper CDP/CLI mode remains available for technical users.

### 6.5 Correction and Challenge. Can patients correct or challenge what the AI produces? **Not applicable to the export, strengthened downstream**

The tool copies records rather than generating judgments, and correcting source records still depends on the institution. But exporting the access log, complete notes, and messages gives patients durable evidence for challenging errors, disputed entries, and record-access questions.

### 6.6 Action Support. Does it help patients understand or act? **Yes**

The export opens with a readable PATIENT_SUMMARY.md and CSV indexes rather than raw JSON alone, and the complete record supports visit preparation, second opinions, advocacy, archiving, caregiver coordination, and any downstream AI analysis the patient selects.

### 6.7 Equity Burden (text finding)

The tool only helps patients whose health systems run Epic MyChart and who have an active portal account, a desktop-class browser, and comfort with bookmarklets or the DevTools console. The interface and outputs are English-only, and no accessibility support is disclosed. Proxy export meaningfully includes caregivers of children and dependents. Patients without digital access, portal literacy, or a desktop environment are left out, and the resulting ZIP concentrates an entire PHI record in one unencrypted local file the patient must store safely.

### 6.8 Data Governance (structured disclosure)

- **Collection:** none by the project. The export runs in the patient's browser tab and writes a ZIP directly to their downloads.
- **Server:** none. The landing page is static and the bookmarklet is self-contained; there is no backend to receive data.
- **Credentials:** never seen, stored, or transmitted. The tool uses the session the patient already holds.
- **CLI mode:** exports land only in local `export*/` directories, which are git-ignored; the README states nothing under them is ever committed or published.
- **Training use, third-party sharing, retention:** structurally not applicable; no party other than the patient touches the data.
- **Verification status:** these claims are checkable in the public source but have not been independently audited, and the hosted bookmarklet build has not been independently verified against the repository.

Overall: local-only by architecture, the strongest data-governance posture in this category. The privacy claim rests on the absence of a server rather than on a policy.

### 6.9 Clinical Boundaries. **Clear**

Clear that it is a data-access tool, not clinical advice. It does not diagnose, triage, or recommend treatment, and no clinical model runs anywhere in the path. The residual risk is downstream: what a patient or their chosen AI tool does with raw records is outside the tool's boundary.

### 6.10 Evaluation Ownership. **Maintainer-defined, source-auditable**

The maintainer defines quality through a public 110-test suite, a synthetic-patient mock instance containing no PHI, and published parity provenance comparing implementations record-for-record, an unusually strong self-verification discipline for an independent project. Public evidence shows no independent security audit, accessibility review, or patient-partnered evaluation, and the repository carries no explicit software license despite the landing page's open-source claim.

## Agency Axis Position: 92

Calibrated against neighbors in the HugoScore directory:

- **Fetch My Epic Token (88):** the closest comparator. Both convert Epic portal access into patient-controlled files, but Fetch My Epic Token routes access tokens and copied session credentials through hosted server-side helpers and demands markedly more technical skill. MyChart Takeout removes the server entirely and lowers the activation burden to a bookmarklet click, so it sits above.
- **OwnChart (94):** a fuller patient-controlled records application. MyChart Takeout is a narrower single-purpose exporter and carries an unstated license, instance fragility, and unaudited claims, so it sits just below.
- **OpenKP (91):** similar source-available, patient-directed posture; MyChart Takeout's serverless data path justifies parity or a slight edge.

The axis position is a lens-orientation cue, not a quality score.

## 7. Key Unknowns

- Whether the hosted bookmarklet build exactly matches the public repository source at any given time.
- Whether the missing software license is an oversight or a deliberate source-available posture; the open-source claim is currently unbacked by a license grant, and the package manifest is marked private.
- Whether replaying internal MyChart APIs creates terms-of-service or account risk for users at particular health systems. HugoScore does not make a legal conclusion.
- How gracefully the exporter degrades across Epic instances with renamed endpoints, disabled features, or different path prefixes, and whether the planned gaps report ships.
- Whether the bookmarklet flow is usable for nontechnical patients, disabled users, low-literacy users, and non-English speakers.
- Whether any independent security review, privacy audit, or accessibility review has occurred.
- How safely patients store and share the resulting ZIP, which concentrates an entire PHI record in one unencrypted local file.

## 8. Patient Agency Interpretation

MyChart Takeout is agency-expanding because it removes the middle layer entirely. Where comparable tools route tokens or session credentials through hosted helpers, this export runs wholly inside the patient's own browser tab and writes only to their own disk. The privacy posture is architectural rather than contractual: there is no server to trust because there is no server.

The second agency pattern is completeness as leverage. Exporting the access log, every message thread, and full clinical notes shifts the patient from portal viewer to evidence holder. A patient can see who accessed their record, preserve notes before they change, and assemble a longitudinal file no single portal session shows them.

The third is deliberate AI-readiness without AI dependence. The deterministic export produces files any AI tool can consume, but the patient decides whether AI enters the picture at all. That separation, plain data path here, patient-chosen intelligence there, is a model CAIHL pattern: the tool builds capacity for critical reflection and strategic action without inserting an intelligence layer of its own.

The constraints are real but comparatively modest. The tool reaches only Epic MyChart patients, relies on internal APIs that vary across instances and could break or raise terms-of-use questions, and assumes a desktop browser and bookmarklet comfort.

## 9. Publication Recommendation

Publishable as an AI-assisted draft profile. Frame as "patient-directed record export and AI-enabling utility," not as a health AI assistant. The 92 axis position reflects best-in-category data governance (no server in the path), explicit no-LLM disclosure, low activation burden, proxy support, and access-log capture, held below the top of the range by the unstated license, instance fragility, unaudited claims, and the baseline portal-access burden. Confidence should remain medium until someone runs a real export, verifies the hosted bookmarklet against source, and the license question is resolved.

## Review Provenance

- Criteria source: HugoScore patient agency framework, CAIHL-derived. Same criteria as every tool in the directory.
- Reviewer: AI-assisted public-source draft prepared in Claude (Cowork agent session). No named human reviewer.
- AI / model: Claude Fable 5 (claude-fable-5), first-pass draft.
- Human review: none. This is an automated draft evaluation and must not be presented as a comprehensively human-reviewed profile.
- Review date: 2026-08-14.
- Limitations: no hands-on export with a real MyChart account, no security audit, no real-PHI code execution, no maintainer interview, no accessibility review, no legal determination. Claims are limited to the project's own public site, repository, and documentation; no independent coverage was located or sought beyond maintainer-background context.

## Change Log

- 2026-08-14: Initial AI-assisted draft evaluation by Claude Fable 5 (claude-fable-5), from a direct request by the project maintainer (Hugo Campos). Axis calibrated against Fetch My Epic Token (88), OpenKP (91), and OwnChart (94).
