A research operating system

Research, from idea to submission.

By hand, or with agents. One workspace carries the whole study — protocol, data, analysis and manuscript — and keeps a record that holds up to review. Agents use the same tools you do, and ask before they decide.

Join the waitlist How it works

Sleep quality and exam stress in medical students — Protocol

Protocol

Cross-sectional surveySTROBElocked · pre-registered

1 missing item — the instrument can’t be final until it’s set.

FieldValueStateVerSource

Stress scalemissingmissing—pack
Primary outcomePSQI global scoreregisteredv1intake
Sample size120amendedv2Methodologist
α0.05registeredv1pack
Recruitment window6–24 Mayregisteredv1you
InclusionYears 1–5, enrolledregisteredv1intake

Latest in log#14Sample size 100 → 120 after pilot drop-outamendment

Illustrative preview — fictional study, not real data.

Reporting guidelines, built in

STROBE
CONSORT
TREND
STARD
SQUIRE 2.0
COSMIN
COREQ
CARE
PRISMA
PRISMA-ScR
RECORD
ARRIVE 2.0
MDAR

A research operating system. One place for the whole study, a record that answers reviewers, and agents that work the same tools you do — always under your supervision.

Complete by hand

Every tool works with no AI at all. A whole paper, start to DOCX, without spending a credit.

Agents, supervised

The same tools, driven by agents you supervise — every change attributed, reviewable and reversible.

An Adversary on your side

A methods reviewer that tries to break your own checks and claims before a journal reviewer does.

From scattered to connected

Six tools. Every hand-off by hand.

A typical study moves between 6–10 disconnected tools — exported, reformatted, re-entered. Why the outcome changed lives in an email thread, and no tool knows which guideline you owe.

One project. Every phase feeds the next.

The same work, connected — with a place to start from an idea, the right guideline built in, and a record of every decision along the way.

?Where do I start?

Idea

?Which guideline?

Guideline

Forms

Survey

Spreadsheet

Data

Stats package

Analysis

Reference manager

References

Word doc

Manuscript

Email thread

Record

Intake

Start from a question. Leave with a design.

Type a rough idea. Oshki asks what a study needs to define, checks PubMed and OpenAlex for what’s already published, and gives an honest verdict — including “not viable as stated” when that’s the truth. Pick a design and your protocol is already seeded.

Intake — start from an idea

What do you want to find out?

Idea
Questions
Prior art
Verdict

Closest published studiesPubMed + OpenAlex

Illustrative preview — fictional study, not real data.

Draft · reason · edit

Every draft shows its work.

Oshki drafts the whole artifact — a questionnaire, an analysis plan, a Methods section — and every element carries one specific reason, linked to your protocol. It opens in the real builder, so you edit one question and keep the rest.

Change the protocol and the reasons that depended on it go stale — flagged, never silently wrong
Methods is written from those reasons, so it says what you collected and why

Survey design — questionnaire

Methodologistdrafted from your Protocol · you’re editingShow all reasons

Q1
How would you rate your sleep quality over the past month?
PSQI · 19 items
Q2
In the last month, how often have you felt nervous and stressed?
PSS-10 · 10 items
Q3
Have you been diagnosed with a sleep disorder?
Single choice
Q4
What year of study are you in?
Single choice

public linkQRMethods will say what you collected — and why.

Illustrative preview — fictional study, not real data.

The research record

A record, not a snapshot.

Protocol fields are versioned, and missing items stay visible. Decisions go into an append-only log with the options you rejected. Every analysis is a recorded run, every check has a verdict against a criterion fixed before or after the data — and every claim links to its evidence.

Corrections supersede; nothing is quietly edited away
Pre-registration as live state, not a PDF
Ran something in SPSS or on the lab server? Log the run by hand

Sleep & exam stress — record

Decision log

append-only

#14
Sample size 100 → 120 after pilot drop-out
Elene · 02 May
amendment
#13
Use PSQI, not a single sleep-hours item
Methodologist · approved by Elene · 28 Apr
decision
#12
Recruit Years 1–5, not 3–5
Ana · 24 Apr
decision
#9
Recruit Years 3–5 only
Elene · 21 Aprsuperseded by #12 →
decision

Checks

Responses reach targetinconclusive
≥ 120 · observed 84pre-data
Primary outcome unchangedpass
PSQI since v1pre-data
Exclude incomplete PSQIpass
< 7 components · 6 removedpost-data

Analysis history

runs

Spearman ρ — PSQI × PSS-10Runner · for Luka1.2 s
Pilot reliability (SPSS)logged by handLuka—

Illustrative preview — fictional study, not real data.

Agents

Agents propose. You decide.

Every tool in Oshki is one operation with two surfaces — the screen you click and a typed tool an agent calls. Agents have exactly your permissions, every change is attributed and reversible, and a new protocol decision always waits for a human yes.

Systematic review — proposals

LibrarianScreens titles & abstracts4 minpending review

Screened 412 records against your inclusion criteria. 38 go on to full text.

Inputs → outputs

Search run #4 · 412 recordsInclusion criteria v2

+214excluded — wrong population
+121excluded — not original research
+39excluded — outcome not measured
→38to full-text review

ApproveEdit & approveRejectDetails ▸

Illustrative preview — fictional study, not real data.

Assisted
one by one
Ask in the rail; the agent calls a tool and you approve each change.
Supervised
batch review
Hand over a multi-step task and review the batch like a pull request.
Headless
scheduled
Long or recurring runs with a budget and a stop rule. Wakes you only to decide.

Seven roles:LibrarianMethodologistRunnerScribeReviewerAdversaryEngineer · code mode

Exports a journal-ready AI-disclosure statement

Analysis

The right test, and the reason for it.

Oshki reads your variable types and design, picks a test you can change, checks its assumptions in plain language, and records the run — inputs, parameters and output — in Analysis history.

Meta-analysis with forest plots, survival and mixed models when the design calls for them
No code, anywhere, unless you switch code mode on

Analysis — Sleep & exam stress

Dataset84 rows

id	psqiord	pss10ord
R-001	9	24
R-002	5	17
R-003	12	29
R-004	4	15
R-005	8	21

Test

Results

Illustrative preview — fictional study, not real data.

Write-up

Every claim, traced to its evidence.

Results are drafted from the runs you recorded and Methods from your protocol. Numbers are references, not retyped literals, so a re-run updates the manuscript. Run STROBE, PRISMA or CONSORT on it any time, and export to DOCX.

Manuscript — Results

Scribedraft · not inserted yet

Of 84 respondents, poorer sleep quality was moderately associated with higher perceived stress during the exam period (Spearman’s ρ = 0.41, 95% CI [0.21, 0.58], p < .001).

AcceptEditDrafted from run #3 — never invented.

Illustrative preview — fictional study, not real data.

Collaboration

One project for the whole team.

Your supervisor, co-authors and statistician work in the same project — not across an email thread and three versions of the same file.

Co-author, supervisor and viewer are presets over real permissions you toggle per person
Decisions route to the person who should make them
Edit the manuscript together in real time; comments and tasks sit next to the work

Write-up — Manuscript

AKLMMT

Results

3 people editing

Approve section

Of 120 students invited, 84 completed the survey (70%). Poorer sleep quality was associated with higher perceived stress (Spearman’s ρ = 0.41, 95% CI [0.21, 0.58], p < .001).

Discussion

Tasks

Enter paper responses, batch 2MTMariam T.
Check STROBE item 13AKAna K.
Rerun correlation on full sampleLMLuka M. · from comment

Illustrative preview — fictional study, not real data.

17 method packs

Pick your design. Get its guideline and a plan you can bend.

Each pack seeds phases, tasks, protocol fields and the reporting checklist your study owes — medicine and health science first, with a generic pack for everything else and a life-sci pack with an opt-in Python notebook. All study types

Quantitative & Experimental

Qualitative & Observational

Evidence Synthesis & Review

Secondary & Computational

Anything else

Quantitative & Experimental

Cross-sectional survey

ChecklistSTROBE

Seed plan · 5 phases

1Protocol
2Survey design
3Data collection
4Analysis
5Write-up

A starting point, not a track — add, reorder, branch and revisit phases as the study demands.

Research, from idea to submission.

Protocol

Complete by hand

Agents, supervised

An Adversary on your side

Six tools. Every hand-off by hand.

One project. Every phase feeds the next.

Start from a question. Leave with a design.

Every draft shows its work.

A record, not a snapshot.

Decision log

Checks

Analysis history

Agents propose. You decide.

Assisted

Supervised

Headless

The right test, and the reason for it.

Every claim, traced to its evidence.

One project for the whole team.

Pick your design. Get its guideline and a plan you can bend.

Be first in when Oshki opens.