Most governance-mature patient AI agent in open source — frontier patterns, working evals, self-hosted, open-core commercial path. Missing explicit AI evidence contract.
Most technically mature entry in batch. Worth monitoring for CLAIM alignment development. Aria commercial path makes it worth engaging for potential PRISM-N or partnership framing.
npm install -g openclaw@latestTula is a skill layer that runs on top of OpenClaw. Install the host runtime first.
openclaw onboardWalks you through gateway, workspace, channels, and LLM keys interactively. Pick a local workspace, not a shared cloud one, to preserve Tula's local-first posture.
git clone https://github.com/realactivity/tula.git && cd tula && pnpm installopenclaw skills add ./Loads the Tula skills, prompts, and FHIR connectors into your OpenClaw workspace. See docs/frontier-agent.md for what each skill does.
openclaw runOpen the local OpenClaw UI in your browser. Tula will be available as an agent. Start with a low-stakes question ("summarize my last lipid panel") before delegating anything consequential.
Apache 2.0 means you can use Tula commercially, but RealActivity also offers Aria as a hosted commercial path — read OPEN_CORE.md to understand the boundary. Tula has working evals (docs/safety-and-disclaimer.md); run them after install to see how the agent behaves on your data shape before trusting outputs.
Curator's guide adapted from the repo's README. Always verify against the upstream source before running. This site does not host or operate any of these tools.
URL: https://github.com/realactivity/tula
Date: 2026-05-28
Analyst: FRAME v1.0 / Synambix
Documents reviewed: README.md, docs/frontier-agent.md (referenced), docs/safety-and-disclaimer.md (referenced), OPEN_CORE.md (referenced)
FOUNDATIONAL | Score: 23/24 | The most governance-mature entry in this batch — frontier agent patterns, working evals, and a safety posture that backs up its patient agency claims.
| Dimension | Score | Evidence |
|---|---|---|
| Patient Agency Index | 3/3 | Self-hosted VM reference deployment; PHI bounded to agent workspace; portal messages drafted, never auto-sent; least-permission skills; patient owns data from first principle |
| Architecture Integrity | 3/3 | OTel-shaped traces; append-only logs; reproducible workspace snapshots; regex secret-scan gates in CI; scope-contained skills; Exchange sender allowlist before model sees email |
| Technical Maturity | 3/3 | Five live skills; Waza compliance in CI; 8/10 tasks passed eval aggregate score 0.97; capability table with Live/In Progress/Planned status; multi-model routing documented |
| Safety & Disclaimer Posture | 3/3 | Dedicated safety-and-disclaimer.md; “not a medical device, not FDA-cleared” explicit; “not a replacement for professional medical advice”; HITL on portal messages documented as architectural constraint |
| Interoperability Stack | 3/3 | SMART on FHIR live; medical PDF and image capture live; EHR portal integration live; multi-backend model routing (Azure, OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, Cohere, open-weight) |
| CLAIM Alignment | 2/3 | Patient agency language is strong and architecturally backed; HITL design is correct; but no explicit citation/evidence contract for AI outputs; no “interrogative stance” tooling visible from available docs; framework is governance-first, literacy-second |
| Sustainability Model | 3/3 | Apache 2.0 open-source; commercial Aria platform as funding path; two named founding contributors; Discussions channel; Contributing guide; enterprise pilot pipeline documented |
| Scope Honesty | 3/3 | Capability table with Live/In Progress/Planned columns; “not a medical device” prominent; open-core split clearly documented; enterprise vs personal use cases separated; roadmap distinguished from product |
| TOTAL | 23/24 |
Contextual Grounding: Strong. Tula grounds all output in the user’s specific health records — SMART on FHIR pulls, PDF uploads, longitudinal memory diffing. The architecture explicitly bounds PHI to the user’s workspace. The system is producing personalized output, not generic health information.
Interrogative Stance: Weak. The current README and skill descriptions do not show mechanisms for patients to question AI outputs, trace claims to sources, or understand what the system does not know. Waza compliance checks correctness from a governance perspective, not from the patient’s epistemic standpoint. A patient using Tula gets governed outputs, not necessarily transparent ones.
Associative Integration: Moderate. Multi-model routing and self-hosted architecture mean the patient could, in principle, connect outputs to other tools they control. But there is no documented path for connecting Tula’s outputs to external knowledge sources or enabling patients to integrate evidence from multiple contexts.
Judgment Layer Activation: Strong by design. HITL constraints mean the patient makes the consequential decisions — portal messages are drafted, not sent; the human confirms before action. This is the right architecture. Whether it builds judgment capacity over time or just adds a confirmation step is not addressed.
Methodological Transfer: Absent. There is no visible mechanism for patients to learn how to reason about their health data through Tula, rather than just receiving its outputs. This may be intentional (the system is focused on coordination, not health literacy), but it represents a gap relative to CLAIM’s full vision.
Tula advances patient agency in the narrow and important sense: the patient’s data stays under their control, AI cannot act consequentially without human confirmation, and the governance architecture is built to resist institutional capture. These are real structural advances over consumer health apps that extract and retain data.
Where Tula falls short of the full patient agency vision is in empowerment depth. The system is built to help patients manage and coordinate their health information efficiently, and it does this with impressive technical rigor. But it does not yet help patients build the reasoning capacity to interrogate their own care — to notice when the AI is confident but wrong, to challenge institutional categories embedded in their records, or to use AI output as a starting point for deeper engagement rather than an end point.
This is not a fatal gap — it may reflect a deliberate staging decision. But it means Tula as documented is a strong coordination tool with governance discipline, not yet a full critical health literacy platform.
None. The README makes no unsubstantiated capability claims. All features are status-labeled. The commercial/open-source split is clearly documented. The Waza eval score is specific and falsifiable.
TRACK + ENGAGE. The most technically mature entry in this batch. Worth monitoring for CLAIM alignment development — if Tula adds an evidence contract and patient-facing interrogative tooling, it becomes the most comprehensive patient AI platform in the open-source space. The commercial Aria path makes it worth engaging for potential partnership framing in PRISM-N or related work.