Full review
OpenKP CAIHL draft report
Evidence-linked HugoScore draft report for a health AI tool that affects patients.
Agency checklist · October 30, 2026
Expands 10 of 12 yes
Document-based draft. Independent confirmation is incomplete. Hugo has not personally verified this service.
Assessed function: Patient-run local MCP server that lets the member's own MCP client (Claude Desktop consumer path; not tested in this review) read, and with explicit confirmation write to, their own Kaiser Permanente NorCal portal record.
Initial Anthropic coding has an Anthropic model in the role record.
A no can mean missing documentation or institution-dependent control; it does not by itself establish harmful behavior. These categories are not clinical safety or product-quality ratings.
- Public documents do not establish actual product behavior.
- Stored quotation missing for yes answers: 5. Evidence repair is pending.
See all twelve answers and their evidence
| Question | Answer | Evidence / review | Source and reasoning |
|---|---|---|---|
| 1. Told it is AI | Yes | documented Confirmed at document level | OpenKP is a local MCP server that lets your AI assistant read, compare, and carefully act on your Kaiser Permanente recordSource The member installs and chooses an AI assistant, and the docs call it AI throughout. There is no separate consent step. |
| 2. Operator and model maker named | Yes | documented Document follow-up: supported | Author: Hugo CamposSource A compromised Anthropic API (Claude Desktop handles this, not us)Source The public design document names Hugo Campos and identifies the Anthropic/Claude connection. Q2 requires public naming, not homepage placement or a single sentence. Applying the same standard as OwnChart changes this answer to yes. Hugo's authorship remains disclosed; it does not justify a stricter naming test. |
| 3. Patient chooses to use it | Yes | documented Confirmed at document level | Want a richer, more agentic interface to their own health record than the kp.org website or Kaiser app provide.Source The README's 'Who this is for' section describes members who install it themselves. Kaiser does not apply it. |
| 4. Patient sets the task | Yes | documented Confirmed at document level | It lets a Kaiser member ask Claude, in plain English, to read and act on their own medical record.Source The tools return raw structured data and do not steer. DESIGN.md says 'Tools don't pre-summarize.' The README's suggested prompts are examples only. |
| 5. Patient can decline or stop | Yes | Self-started rule; direct refusal wording not documented Confirmed at document level | Quotation not recorded; evidence repair pending. Source Yes follows the adopted Q5 rule for a self-started optional tool, rather than an explicit published refusal right. The original coder's note below predates that normalization and is retained as history: No source states that the patient can decline or stop and keep the same service. Related but not the same statements: DESIGN.md says 'Everything a member can do by clicking around kp.org, OpenKP exposes as a tool', and 'If OpenKP disappears tomorrow, the user loses nothing' (about credentials). Write actions need a preview and explicit confirm. I did not treat these as an explicit decline right. |
| 6. Patient sees the AI output | Yes | documented Confirmed at document level | Claude shows this to the user.Source The statement is about write previews. Read answers appear in the patient's own client, as in the openkp.org examples. Preview and commit actions are logged in ~/.openkp/audit.log. |
| 7. Patient can correct or challenge | No | not-documented Document follow-up: not documented in inspected sources | returns a preview without touching KaiserSource GitHub access is now available. The root README, design document and installation README were inspected. They document preview and confirmation of writes, but no route to correct or contest the assistant's output was found in those documents. No retained under the not-documented rule. This is a bounded document finding, not proof that conversational correction is impossible. |
| 8. Patient can get a copy | Yes | documented Confirmed at document level | Data exports include conversation data and the user data for your account.Source Individual-user Claude export provides a qualifying copy; deletion of all local or institutional artifacts is not established. |
| 9. Patient can delete | Yes | documented Confirmed at document level | which will be removed immediately from your conversation history and automatically deleted from our back-end within 30 days.Source Host terms (Claude Desktop, Anthropic). OpenKP documents no deletion or uninstall steps for its own local files in ~/.openkp. Its audit log is described as append-only. Deleting local files is possible but not documented. |
| 10. Training or sale does not override patient choice | Yes | documented Confirmed at document level | We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings.Source Host terms decide, for the documented default of Claude Desktop on a consumer account. Opt-out exists, but the policy still allows use of data flagged for safety review or explicitly reported. Conflict in OpenKP's own docs: the README says 'PHI passes only between Kaiser, OpenKP on your Mac, and Claude Desktop on your Mac' and 'Everything stays on your machine.' The website's Limits section is more accurate: 'Your assistant sees tool results'. Claude Desktop sends tool results to Anthropic's cloud. A different MCP client would change this answer. |
| 11. Patient can run it or take it elsewhere | Yes | documented Confirmed at document level | Source-available. Noncommercial license (PolyForm NC 1.0.0). Local-first.Source The repo is public and openkp/LICENSE is PolyForm Noncommercial 1.0.0 (Required Notice: Copyright (c) 2026 Hugo Campos). The README says 'Everything stays on your machine. There is no OpenKP server' and the site says any MCP-capable client should be able to call it. The patient runs it locally with their own Kaiser credentials. It is source-available, not OSI open source. It covers Kaiser NorCal only. The site warns that automated portal access 'may conflict with Kaiser's terms of use'. |
| 12. No advertising or data business | No | documented Corrected after independent review; accepted by editor | sharing your personal data for targeted advertising to promote our products and servicesSource Claude consumer targeted advertising fails Q12; training alone is a separate Q10 issue. |
Initial coding: Claude Sonnet helpers, compiled by Claude Opus 5.5 (Anthropic). Selective review and corrections: GPT-6 Astra (OpenAI). Editor: Hugo Campos. Acceptance of corrections is not product verification.
Checklist rules · Independent review and limitations · Later evidence follow-up
Earlier report
OpenKP: CAIHL reassessment
September 8, 2026. Published AI-assisted draft.
This assessment covers OpenKP with a patient-chosen AI client. OpenKP itself is a local record connector, with preview and confirmation workflows. The chosen AI client receives returned records under its own policies. OpenKP and HugoScore share a maintainer, so this is a self-evaluation.
Scope and agency posture
OpenKP plus a patient-chosen MCP AI client, for Kaiser Permanente Northern California records. OpenKP itself is a local connector, not a language model. HugoScore and OpenKP share a maintainer, so this remains a disclosed self-evaluation.
Posture: Strongly agency-expanding.
Axis: 91/100, retained provisionally. No numerical recalibration was performed.
Sources and documented findings
- The website describes patient questions over portal records, explicit write confirmation, and a separate AI client that receives tool results.
- The current repository README documents preview/commit calls and a noncommercial license permitting personal and advocacy use. An older retrieved package README still says MIT and overstates local-only data handling.
- The retrieved server source exposes record-retrieval tools. Its cached version was older than the website, so it does not establish current runtime parity.
Patient authority
Inference from the documented workflow: The patient chooses the substantive questions and AI client, and operates the portal connector. A callable confirmation parameter is a control mechanism, but not proof that every client obtains informed human approval before invoking it.
Critical capacity
Editorial assessment: Cross-note comparisons, access-log questions, and retrieving original records enable scrutiny of institutional accounts. The language model provides interpretation, so accuracy and critical capacity must be evaluated for the combined workflow.
Informed control
Editorial assessment: The website identifies the client data boundary more clearly than the repository's broad local-only statement. A source-available noncommercial license supports personal adaptation without guaranteeing universal use rights or secure operation.
Assessment and limits
The strong design-based finding is supported, but the score should explicitly describe the assembled AI workflow. Apply the same distinction to deterministic exporters later. No demonstrated outcome or independence claim is warranted.
Confidence: Medium for architecture, lower for current implementation and usable approval.
Remaining uncertainty: Client approval behavior, current source/version parity, external model data handling, portal terms, and accessibility outside technically capable NorCal users.
Published documentation is evidence of stated conditions, not proof of actual implementation. Unknowns did not receive automatic negative points. Funding, sponsorship, and public code do not determine agency by themselves.
Review provenance
- Reviewer/model: OpenAI Codex / GPT-6.
- Method: Focused public-source reassessment using CAIHL: patient authority, critical capacity, and informed control. Existing evidence plus one focused primary-source pass and at most one targeted follow-up. No live product testing. Numerical scores remain provisional editorial placements, not a new calculation.
- Human review: Hugo Campos authorized publication of these AI-assisted draft reassessments on September 8, 2026. This does not claim comprehensive human verification of every finding.
- Earlier review: 2026-06-08. Historical assessment. Earlier claims are not automatically reverified by this publication.
Disclosures
Self-evaluated: OpenKP and HugoScore share the same maintainer. Hugo Campos maintains both OpenKP and HugoScore, so this profile is a self-evaluation rather than an independent review. It is published with that label, and independent third-party review of OpenKP is invited and outstanding.