An AI research agent for real medical studies, built to turn data, evidence, and drafts into manuscript-ready work
Disease Studies · Evidence Building · Analysis Support · Manuscript Delivery
|
Who It Serves Clinicians, PIs, and medical research teams working with disease-specific data and moving studies toward manuscripts |
What It Organizes Study questions, data assets, analysis progress, evidence gaps, and manuscript-facing files inside one governed workspace |
How To Start Tell it the disease area, the dataset you have, the question you want to answer, and the paper outcome you want |
Med Auto Scienceis for teams already doing real medical research. It keeps study questions, data, analyses, evidence, drafts, and delivery files connected on one governed study line so the work can keep moving and stay reviewable.
The hard part of medical research is rarely a single paragraph. It is moving a study from data and ideas toward a paper that can be reviewed, revised, and submitted.
Medical teams often run into the same problems:
- You have data, but it is not clear which question is worth pursuing.
- You have preliminary results, but they do not yet form one manuscript line.
- Figures, drafts, analysis notes, and validation results drift across folders and conversations.
- Review, extra analysis, claim revision, and delivery files overlap until progress is hard to explain.
- Multiple disease studies run in parallel, and key evidence or decisions can be lost.
Med Auto Science is built around those research problems.
It organizes a medical study as a research line: frame the valuable question, prepare the data and evidence, advance analysis and validation, build the manuscript story, and keep paper-facing files ready for review and delivery.
You can start with prompts like:
- "Help me find a paper-worthy question from this colorectal cancer dataset, tell me what evidence is still missing, and propose the next step."
- "I already have preliminary results. Turn them into one manuscript line and tell me what validation to do next."
- "Keep pushing this disease study toward a publishable paper, and keep the progress plus files organized as we go."
|
Identify paper-worthy questions from data Start with a disease cohort, registry, or real-world dataset, then identify questions with clinical value, evidence support, and manuscript potential before piling up analyses. |
Turn scattered results into a manuscript line Existing analyses, early findings, figures, and drafts are organized into a clearer research story with explicit next evidence steps. |
|
Keep progress and delivery files together Tasks, files, figures, drafts, validation notes, and deliverables stay tied to the same study workspace so the line remains reviewable and easy to continue. |
Let AI do the heavy lifting while researchers keep judgment AI can help prepare data, run analyses, organize evidence, and report progress. Clinical framing, claim acceptance, and final submission decisions stay with researchers and PIs. |
- Finding a study question worth continuing from a disease-specific dataset, registry, or cohort.
- Turning existing analyses and early results into one manuscript line.
- Managing validation, subgroup analysis, calibration, clinical utility analysis, and other supporting evidence.
- Keeping multiple related studies organized in one workspace.
- Keeping paper-facing results, figures, drafts, and delivery files tied to their study.
Med Auto Scienceis publicly positioned as a medical researchFoundry Agent: an independent domain agent for turning disease data, study questions, evidence, and manuscript work into one governed research line.- MAS is also an
OPL-compatible package built on OPL Framework. OPL can discover MAS stage descriptors, action metadata, handoff contracts, receipts, and projections; MAS remains the medical research owner. - MAS now keeps its canonical OPL Agent semantic pack under repo-root
agent/and declares those required files through repo-rootcontracts/*.json. OPL compiles that pack into the generated CLI / MCP / Skill / product-entry / tool descriptors. MAS repo-owned code is limited to medical authority functions, owner receipt / typed blocker production, domain handler targets, refs-only projections, and necessary medical helper implementation. - MAS keeps the direct app skill path as a first-class entry. Direct MAS activation and OPL-hosted handoff return to the same MAS-owned stage, controller, durable truth, review, and artifact surfaces; when a task is started in the hosted path, persistent wakeup, retry, resume, and attempt scheduling are enabled by default through OPL/Temporal rather than a Codex App outer driver.
- MAS owns medical research truth, quality verdicts, runtime-facing owner receipts, artifact authority, and publication authority. OPL Framework owns the generic runtime platform; framework metadata does not replace MAS owner surfaces.
- AI-first quality gates require separation of labor: an execution agent such as Codex CLI may produce the stage work, but reviewer/auditor judgment must be invoked as an independent agent task with its own context, task record, and receipt. A single agent cannot execute and then review itself to close MAS quality gates.
- This positioning adds no MAS-owned daemon, scheduler, or attempt loop. Hermes-Agent, MedDeepScientist/MDS, and the retired MAS local scheduler remain explicit optional/provenance/tombstone surfaces, not default public targets.
- Manuscript quality is governed by MAS study charters, evidence ledgers, review ledgers, AI reviewer workflow, publication gates, and controller records. Status panels, script checks, and historical MDS coverage provide supporting evidence.
- Clinical framing, claim acceptance, and final submission decisions stay with researchers and PIs.
- Journal submission and external system interaction stay under human supervision.
Technical boundary for operators
Med Auto Scienceis a medical research domain agent and Foundry Agent. It can be called directly by Codex, and it can also be discovered and hosted as an OPL-compatible package underOPL Framework.- MAS owns the medical work itself: study intake, workspace context, evidence progression, progress explanation, manuscript quality judgment, runtime-facing owner receipts/projections, artifact authority, and manuscript-facing delivery.
OPL Frameworkis the upper stage-led framework. It owns the generic runtime platform: stage attempts, queues, wakeups, recovery, approvals, receipts, state-machine execution, and cross-domain projection. MAS keeps medical conclusions, manuscript quality, domain transition semantics, artifact authority, and submission-facing judgment.- In the OPL framework, a
Stageis a large task step such as scouting, analysis, writing, reviewer repair, or delivery. An Agent executor is the minimum execution unit inside a stage;Codex CLIis the current first-class executor. - MAS has completed monolith closeout.
MedDeepScientist/DeepScientistremains available as provenance, explicit archive import, backend audit, upstream learning, and parity reference. - Long-running OPL-hosted production execution is Temporal-backed. Temporal is the required production provider for OPL durable stage attempts, signal/query, retry/dead-letter, and workflow history.
Hermes-Agentis not the target session/wakeup substrate, but it remains available as an explicit Agent executor adapter / proof lane that promises connectivity and auditability, not behavior or quality equivalence withCodex CLI.
- Potential users and medical experts should start here, then continue to the Docs Guide.
- Technical readers and planners should read Project, Status, Architecture, Invariants, and Decisions.
- Developers and maintainers should continue from the Docs Guide into
docs/active/,docs/runtime/,docs/delivery/,docs/references/, anddocs/policies/.
Start here if you are handing this repo to Codex or another agent
- No. Cloning this repo does not auto-install OPL Framework or the production runtime. To make MAS usable, first make the current
one-person-labcheckout or release bundle available, then usemedautosci product skill-catalog --profile <profile> --format jsonto inspect the skill ormedautosci domain-handler export --profile <profile> --format jsonto hand off to OPL or Codex. - Read the Docs Guide first. It maps the current product boundary, operator entry surfaces, and the technical reading order.
- If you need to bootstrap or take over a disease workspace, read Bootstrap next. It explains the workspace-first model and the
init-workspace -> doctor -> show-profile -> bootstrappath. - Treat Project, Status, Architecture, Invariants, and Decisions as the repo-tracked human-readable truth set before changing runtime or docs.
- The current operator entry surfaces are discoverable as
CLI,MCP,product-entry, andcontroller, but their generic descriptors and hosted shells are owned by OPL generated/hosted surfaces. The repo-rootagent/pack is the generated-interface semantic source for OPL; local entry code only serves MAS domain handler targets, medical authority functions, owner receipt / typed blocker producers, or refs-only projections. - MAS can be invoked directly through its Codex app skill or through OPL. Both routes use the same MAS-owned stage, controller, durable truth, and artifact surfaces; OPL/Temporal is the default hosted autonomous runtime for durable scheduling, wakeup, retry, resume, and projection.
- New disease workspaces are no-root-Git / no-quest-Git by default. OPL owns runtime lifecycle, provider attempts, wakeup, retry, and resume state; MAS only exposes domain refs, restore/provenance locators, artifact authority, owner receipts, typed blockers, and human gates.
- When an external agent needs the repo-tracked MAS skill surface directly, use
medautosci product skill-catalog --profile <profile> --format json; it returns the single MAS app skill, the underlying command contracts, and body-free owner-route / artifact / authority refs for OPL or direct Codex execution. - For OPL Full online handoff,
medautosci domain-handler export --profile <profile> --format jsonexposes body-free owner-route refs andmedautosci domain-handler dispatch --task <task.json> --format jsonrecords MAS owner-route dispatch receipts. OPL owns stage graph hydration, queue, attempt ledger, retry/dead-letter, and provider readiness; MAS owner surfaces only return receipts, typed blockers, human gates, or domain refs.