Copilots · The platform
Build AI apps without handing control to the AI. The AI proposes. Your team decides.
For teams that run AI apps for customers or staff and need to know exactly who changed what, and why.
See a reviewed change- Requested›
- Reviewed›
- Live
- "max_tokens": 5000+ "max_tokens": 6000
Every change to an app arrives as a request with a diff and a risk score. Someone on your team approves or rejects it with a reason, and only then does it reach the stage you picked. Every version is kept and can be rolled back. You build the app by describing it to the Architect, which writes those requests for you. Sign-up, roles, agent routing, data masking, processes and cost limits are already part of the platform, and a new app, a new stage or a release never needs a server deployment.
- AI suggests, people decide.
- Protected stages are locked in the database.
- Personal data stays masked by default.
The results on this page come from the platform's own code or were recorded in the example app, and each block is labelled. Inputs were frozen on 1 Nov 2026, which is why the page works without an account and makes no AI calls. All people, practices, tariffs and invoices are made up.
Review and rollback
Every change is reviewed, kept and can be rolled back
Nothing the AI proposes goes live on its own. Each change shows exactly what it alters, gets a risk score and passes a quality gate. A person approves or rejects it and the reason is stored with it. Apps live inside the platform, so a new app, a new stage or a release is a reviewed configuration change. No server deployment is involved.
Review this change yourselfcomputed by the platform
- Requested›
- Reviewed›
- Live
{"id": "pro","label": "Pro","features": ["chat_basic","claim_check","image_analysis"],- "max_tokens": 5000+ "max_tokens": 6000}
Restoring version 12 of example-app/testexamplecomputed by the platform
- ↺ profile_schema
- ↺ tier_config
- ↺ app_config
Stages of the example appfrom the example app
Ten test stages are as fine as one. Protected stages can only be changed through one write path that asks for a reason and records the before and after. A database trigger blocks everything else. Staging compares two stages entry by entry and moves sections across, and a hotfix needs a written reason. To keep reviews short, related requests travel as one package, and an app can let the Architect write straight into its open development stage. Protected stages always need a review.
Backups
The Architect
How you work with the Architect
The Architect is a chat in the admin area. It knows the current configuration of your app and the rules of the platform. It does not only build what you ask for: it checks the app, compares parts that belong together, suggests what your app could do next and builds it for your review. It never changes anything on its own. Whatever it proposes becomes a change request that a person reviews.
- Step 1System checkWhen you open it, the Architect looks through the app's configuration and suggests a few next steps.
- Step 2Your promptYou describe what you want in plain words.
- Step 3Answer and previewIt replies with a plan and shows what would change: a field, a role, a process, a widget, a schedule.
- Step 4Change requestEach change comes as a card with the area, a risk score and a short risk analysis. You submit it with one click, or several at once as a package.
- Step 5ReviewSomeone checks the diff and the quality gate on the review board and approves or rejects it with a reason. A snapshot is taken before anything is applied.
It checks, suggests and builds for review: one recorded sessionfrom the example app
When the chat opens, the Architect reads the configuration of the app and stage and writes a health check.
- 10 agents, routing rules complete
- 4 business objects, claims and medical invoices with a hand-over to case handlers
- 2 processes: claims triage (by hand) and storm precautions (daily 07:00)
- A dashboard tile "metric" has no widget and stays empty
- The admin dashboard has no KPIs of its own
Recorded in the example app on 2 Nov 2026, translated and shortened, no live AI calls. Findings are proposals that a person checks: the same check also reported invoice rules as missing that exist, and called all four objects handed over when two are. Those are left out or corrected here.
Built in one day
One insurance app. Built in one day.
Atomar builds AI-powered business applications. It connects data, automates workflows and brings customers and staff into the same process, with every change reviewed and versioned. The example: claims and invoices handled end to end, a mailbox for customers and a team inbox for case handlers, automated storm precautions and interactive apps. All of it configured on Atomar by talking to the Architect, with every change reviewed. Nine requests were rejected with a reason and fixed in the next one. That is exactly what the review is for.
- Create the app in the App Registry
- The master template is copied into every stage
- Admin rights and a workspace are set up
- 19 configurations per stage, 26 AI modules
- 1Prompt
- 2Request with risk score
- 3Review
- 4Live in the chosen stage
What the day producedfrom the example app
- Claims with photo, handed over to case handlers with consent, the status flows back to the customer
- Medical invoices: the AI reads the PDF, tariff rules work out the refund, a case handler takes it from there
- A team inbox for case handlers and a damage estimator for customers
- Storm warning: a daily process asks a weather API and creates a task when a storm is coming
- Cover check across insurance lines, contract cockpit with contracts, lines and claims
- Mailbox for documents from the insurer
- Nutrient detective: a learning game with three modes, knowledge cards and a saved high score
- Widgets are small apps in a sandbox, not only dashboard tiles
Also in the app: 5 KPIs, 4 widgets, 2 processes, 4 specialist agents with a model council, onboarding with legal consent, tiers with daily limits for widget AI and invoice checks.
Here are 25 of the 50 requests from 17 configuration areas, in order. We rewrote them without customer details. Prompts were not stored yet when this app was built, so what you see is what the Architect made of them. Today every prompt is kept in the audit log.from the example app
| Time | Area | Request | Risk | Result |
|---|---|---|---|---|
| 0:00 h | app_config | App name and tagline, metrics and objects switched on. | 5 | approved |
| 0:01 h | app_roles | Three roles: policyholder, family member, advisor (may invite). | 5 | approved |
| 0:01 h | profile_schema | Four profile fields: postcode, household size, existing cover, contact preference.Rejected: A multi-select was typed as a single select. The next request fixes it. | 8 | rejected |
| 0:03 h | profile_schema | Profile fields corrected (multi-select), template placeholders removed. | 18 | approved |
| 0:03 h | object_registry | Objects "insurance contract" and "claim".Rejected: AI field descriptions under a key the runtime never reads. | 10 | rejected |
| 0:07 h | object_registry | Objects corrected. Same content, right key. | 10 | approved |
| 0:07 h | profile_schema | Surname, e-mail and postcode masked: the AI only sees placeholders. | 22 | approved |
| 0:07 h | routing_capabilities | Four specialist agents for the chat.Rejected: Sent back in review and resubmitted as a purely additive change. | 15 | rejected |
| 0:14 h | routing_capabilities | Four specialist agents: claims, cover & benefits, contract changes, emergencies. | 5 | approved |
| 0:14 h | onboarding_legal_content | Three legal notices for onboarding (not advice, AI notice, privacy). | 8 | approved |
| 0:14 h | metric_definitions | Three KPIs: open claims, next due date, cover check. | 10 | approved |
| 1:24 h | object_registry | Photo field on the claim object, sent as a delta only. | 5 | approved |
| 1:27 h | processes | Process "claim triage" with 5 nodes.Rejected: Incomplete edges (a node without a successor). | 8 | rejected |
| 1:29 h | processes | Claim triage with 6 nodes. It loads the claims, lets the AI check them for missing details and creates a task for each incomplete one. | 8 | approved |
| 1:45 h | rat_triggers | Keywords on which the model council (several models) decides.Rejected: Would have replaced the existing word list instead of extending it. | 5 | rejected |
| 2:08 h | api_connectors | Weather service connector that turns a postcode into coordinates.Rejected: Quality gate: writes profile fields that do not exist. | 5 | rejected |
| 2:16 h | api_connectors | Two connectors to a public weather service. No key needed and no writes to the profile. | 5 | approved |
| 2:16 h | processes | Process "storm precautions". Every day at 07:00 it checks the forecast and creates a task when a storm is coming. | 8 | approved |
| 2:42 h | object_registry | Hand a claim over to case handlers with the customer's consent. | 8 | approved |
| 3:25 h | ai_strategy | Routing logic for all agents (which model, which context). | 42 | approved |
| 5:41 h | custom_widgets | Three widgets: cover check, claim estimator (with AI), contract cockpit. | 15 | approved |
| next day 09:52 | claim_rules | Benefits catalogue for a made-up tariff: rules for medical fee scales, self-pay services and more. | 12 | approved |
| next day 09:52 | object_registry | Medical invoice object: PDF, invoice check, result and hand-over to case handlers. | 12 | approved |
| next day 10:01 | object_registry | Mailbox for documents from the insurer, with a KPI for unread ones. | 6 | approved |
| next day 10:28 | custom_widgets | Prevention game for customers, "nutrient detective": three modes, knowledge cards, high score in the profile. | 8 | approved |
Under the hood
One platform behind the whole application
The Architect configures connected building blocks on an existing platform. Screens are only one of them. Your app's data, interfaces, AI agents and workflows work together, with reviewed and versioned changes.
Processes and work inbox
Build business processes by pointing and clicking
Automate end to end, and keep people in the loop where you choose. A process connects data, API calls, AI steps and branches, and adds tasks for people wherever your workflow needs them. Without a task it runs on its own from start to finish. You draw it in the visual editor or describe it to the Architect, and it is reviewed like every other change. When a step needs a person, the process creates a task in that person's task list and waits. Once the task is done, it carries on with everything it has gathered so far. Cases handed over to a team land in the team inbox of the case handlers. A scheduler runs timed processes and a monitor shows every run.
Claims triagefrom the example app
- 1triggerStart by hand
- 2object queryLoad all claims
- 3ai logicAI checks each open claim for missing details and urgency
- 4conditionAny incomplete claims?yes: next step, no: done
- 5human taskComplete the claim: {{item.title}}one task per incomplete claim
- 6endTriage done
Drawn from the example app's stored process configuration as of 1 Nov 2026, labels translated. It shows the structure, not a live run.
The task in the person's task listexample
Team inbox: a claim from the customer to the case handlers and backfrom the example app
- CustomerHands the claim over and agrees. Only the shared fields go across.
- Case handlerTakes the case from the team inbox.
- Case handlerAsks the customer a question.
- CustomerAnswers, and the case goes back to the case handler.
- Case handlerCloses the case with a comment.
Storm precautions: the condition, evaluated by the process enginecomputed by the platform
- triggerDaily at 07:00 (cron 0 7 * * *)
- api_callPostcode to coordinates (weather service)
- api_callFetch the forecast for the next days
- conditionStorm threshold exceeded?
- human_taskTask: secure the basement, keep photos ready for a claim
- endDone
Math.max(...forecast.data.daily.wind_gusts_10m_max) > 70 || Math.max(...forecast.data.daily.precipitation_sum) > 30
API connectors
Connect APIs visually
A connector is an endpoint with placeholders. In the editor you set the address, which values fill the placeholders, what comes back and a preview. A placeholder is filled from the process step before it, from fields you bind by hand, or from the profile field of the same name. Keys are stored on the server and never reach the browser. The storm process uses two of them, one after the other.
https://geocoding-api.open-meteo.com/v1/search?name={{plz}}&country=DE&count=1https://geocoding-api.open-meteo.com/v1/search?name=44139&country=DE&count=1
https://api.open-meteo.com/v1/forecast?latitude={{lat}}&longitude={{lon}}&daily=precipitation_sum,wind_gusts_10m_max&forecast_days=2&timezone=Europe/Berlinhttps://api.open-meteo.com/v1/forecast?latitude=51.51&longitude=7.47&daily=precipitation_sum,wind_gusts_10m_max&forecast_days=2&timezone=Europe/Berlin
The addresses are filled by the platform's own code for a made-up profile with the postcode 44139 and a frozen answer of the first service. A placeholder that cannot be filled stops the call instead of sending it half empty.
In the app
What your users get
Every app built on the platform comes with the same set of building blocks. The Architect decides which ones an app uses, what they are called and how they look. Sharing works across roles, so the same blocks fit very different teams. A trainer and their clients share exercises, notes and a coaching diary. In a company, a dispatcher hands jobs to technicians, and both see the notes from the site.
Users and onboarding
Sign-up, roles and teams come from configuration
Onboarding is a list of steps in the app's configuration. The Architect writes it like everything else, and each user goes through it once per app and stage.
Onboarding of the example appfrom the example app
- 1 · authYour accountSign in with Google
- 2 · rolesYour roleOnly roles marked self-selectable
- 3 · namesYour profileFields come from the profile schema
- 4 · legalNoticesNot advice, AI notice, privacy. Consent is recorded.
- 5 · teamJoin a teamInvite link with a role
Roles and who may invite whomfrom the example app
| Role | Kind | May invite |
|---|---|---|
| Policyholder | customer | family members |
| Family member | customer | nobody |
| Advisor | customer | policyholders, family members |
| Case handler | staff | nobody, assigned by an admin |
Access per app and stage
Privacy by default
What the model sees and what an admin sees
Fields marked as masked in the profile schema never reach the model as real values. An agent can get a documented exception for a single field. Admins see the same data masked until someone with the right permission unmasks it and gives a reason.
The user writesexample
Hi, this is Max Muster from 44139 (max.muster@example.org). Is my home contents cover enough for a household of 3?
The model receivescomputed by the platform
Hi, this is Max [MASKED_LASTNAME] from [MASKED_PLZ] ([MASKED_EMAIL]). Is my home contents cover enough for a household of 3?
Profile context for the modelcomputed by the platform
PROFILKONTEXT - Aktuelle Rolle des Users: Policyholder (System-ID: policyholder) - First name: Max - Surname: [MASKED_LASTNAME] - E-mail: [MASKED_EMAIL] - Postcode: [MASKED_PLZ] (KI-Hinweis: Postcode of residence, relevant for regional services.) - Household size: 3
The headings of this block are German because the platform's prompts are. The values come from the app.
The answer, translated back for the usercomputed by the platform
Hi Max, for a household of 3 in [MASKED_PLZ] it is worth checking the sum insured. I will send the details to [MASKED_EMAIL]. Hi Max, for a household of 3 in 44139 it is worth checking the sum insured. I will send the details to max.muster@example.org.
The answer itself is a fixed example. Masking and unmasking run the chat's own code.
The admin user listexample
| Name | Role | Tier | |
|---|---|---|---|
| Policyholder | Pro | ||
| Family member | Free |
Only global admins can unmask. They have to give a written reason, the action is recorded in the audit log, and the data is masked again after 5 minutes.
Emergency keywords are handled before any AIcomputed by the platform
Chat and agents
Each question goes to the agent that should answer it
A router picks a specialist agent for every question. Each agent has its own prompt, persona, model and data rights. Only the reimbursement agent sees medical invoices, for example. Models from several providers can be switched on or off per app and stage. Sensitive questions can go to a council of models. It only runs on configured trigger words or when a user picks it, because several models take longer and cost more, and every call counts against the daily limit of the user's tier. Each platform function, from the Architect to the router, can use its own model, including your own through an OpenAI-compatible endpoint.
Agents of the example appfrom the example app
| Agent | Purpose | Model | May see |
|---|---|---|---|
| general_support | Default | Claude | selected profile fields, contracts |
| claims | Report a claim, checklists per line | Claude | claims, contracts |
| emergency | Acute emergencies | Claude | claims |
| cover | What am I covered for? | Gemini | contracts |
| contract_changes | Moving, changes | Claude | contracts, postcode (unmasked for this agent only) |
| medical_invoices | Invoices, reimbursement | Claude | medical invoices, contracts |
| image_analysis | Photos and documents | Claude | nothing extra |
Inside one agent: what it knows and what it may seefrom the example app
- Never promise payments, settlements or contract changes. Only the insurer decides.
- Ask only for what the next step needs. Never ask for diagnoses, policy numbers, bank details or ID in the chat.
- If someone is in danger, the first answer is the emergency number 112.
- Then: insurance line and date, only the details that line needs, a summary to report it.
Routed in the live testfrom the example app
Reviewed knowledge refinement
The model council's answers can be distilled into new knowledge or new routing rules. The result becomes a change request on the review board, never a direct write. If you choose, it also lands as memory candidates, and a candidate is only used once a person approves it.
Rules, not guesses
Invoice check: the AI reads, the rules decide
In the app, a model only reads the positions off the invoice PDF. Rules from the configuration decide whether and how much is paid back, and every reduction shows how it was calculated. Here the reading is frozen and the rules run live. The example app is German, so rules and results appear the way its users see them.
The tariff's rules, where the first match appliesfrom the example app
| Rule | Matches | Paid back |
|---|---|---|
| Optional self-pay services (IGeL) | "igel", "auf wunsch", "kosmetisch" | not covered |
| Lab services (GOÄ section M) | GOÄ · 3500 to 4787 | 100 % up to factor 1.15 |
| Technical services (GOÄ sections E, O) | GOÄ · 500 to 569, 5000 to 5855 | 100 % up to factor 1.8 |
| Medical services (GOÄ) | GOÄ | 100 % up to factor 2.3 |
| Non-medical practitioners | GebüH | not covered |
| Glasses and lenses | "brille", "kontaktlinse" | 80 %, at most €150.00 per invoice |
| 1 | Beratung, auch telefonischÄrztliche Leistungen (GOÄ): vollständig erstattungsfähig (Faktor 2,3 im Rahmen). | €10.72 | €10.72paid |
| 5 | Symptombezogene UntersuchungÄrztliche Leistungen (GOÄ): vollständig erstattungsfähig (Faktor 2,3 im Rahmen). | €10.72 | €10.72paid |
| 250 | Blutentnahme aus der VeneÄrztliche Leistungen (GOÄ): vollständig erstattungsfähig (Faktor 2,3 im Rahmen). | €5.36 | €5.36paid |
| 3550 | Blutbild (Laborleistung)Laborleistungen (GOÄ Abschnitt M): vollständig erstattungsfähig (Faktor 1,15 im Rahmen). | €4.02 | €4.02paid |
| 1 | BeratungAbgerechnet mit Faktor 3,5. Der Tarif erstattet Ärztliche Leistungen (GOÄ) bis Faktor 2,3: 16,32 € × 2,3 / 3,5 = 10,72 €. | €16.32 | €10.72reduced |
| 7 | Vollständige Untersuchung BewegungsorganeAbgerechnet mit Faktor 3,5. Der Tarif erstattet Ärztliche Leistungen (GOÄ) bis Faktor 2,3: 32,64 € × 2,3 / 3,5 = 21,45 €. | €32.64 | €21.45reduced |
| 5135 | Röntgen Brustorgane, ÜbersichtAbgerechnet mit Faktor 2,5. Der Tarif erstattet Technische Leistungen (GOÄ Abschnitte E, O) bis Faktor 1,8: 40,80 € × 1,8 / 2,5 = 29,38 €. | €40.80 | €29.38reduced |
| 3550 | Blutbild (Laborleistung)Laborleistungen (GOÄ Abschnitt M): vollständig erstattungsfähig (Faktor 1,15 im Rahmen). | €4.02 | €4.02paid |
| 1 | Eingehende UntersuchungBehandlungen durch Heilpraktiker sind im Tarif nicht versichert. | €20.50 | €0.00not covered |
| 20.1 | AkupunkturBehandlungen durch Heilpraktiker sind im Tarif nicht versichert. | €45.00 | €0.00not covered |
| 25.4 | SchröpfenBehandlungen durch Heilpraktiker sind im Tarif nicht versichert. | €18.00 | €0.00not covered |
| 1 | BeratungÄrztliche Leistungen (GOÄ): vollständig erstattungsfähig (Faktor 2,3 im Rahmen). | €10.72 | €10.72paid |
| 5 | Symptombezogene UntersuchungAbgerechnet mit Faktor 3,5. Der Tarif erstattet Ärztliche Leistungen (GOÄ) bis Faktor 2,3: 16,32 € × 2,3 / 3,5 = 10,72 €. | €16.32 | €10.72reduced |
| A750 | Hautkrebs-Screening auf Wunsch (IGeL)Individuelle Gesundheitsleistungen ohne medizinische Notwendigkeit sind im Tarif nicht versichert. | €35.00 | €0.00not covered |
| · | Brillengläser Gleitsicht, 2 StückErstattungssatz 80 % von 380,00 € = 304,00 €. Höchstbetrag für Sehhilfen: 150,00 € je Rechnung, davon noch 150,00 € offen. | €380.00 | €150.00reduced |
| · | Brillenfassung TitanErstattungssatz 80 % von 120,00 € = 96,00 €. Höchstbetrag für Sehhilfen (150,00 € je Rechnung) ist mit den vorigen Positionen ausgeschöpft. | €120.00 | €0.00not covered |
Widgets
Mini-apps in a sandbox, also built by prompt
More than dashboard tiles: interactive mini-apps with their own interface, animation, keyboard controls and sound. They run on their own or connect to the app's data and to AI on the server through declared inputs, events and actions. Widgets run in a sealed sandbox. They can't touch the page or its cookies, they may load libraries from four approved code CDNs and reach no other server, and they only exchange messages with the app through declared inputs and events. A widget can call an AI agent on the server, and users can build their own by prompt within the limits of their tier. Every widget below runs live in that same sandbox.
Insurance, connected to the app. This widget gets the customer's policy from the profile through declared inputs and reports back through declared events. The panel next to it shows what the platform passed in and every event the app received after validation. Pick a scenario and start a claim.
- customer_nameSam Keller
- policy_tierComfort
- deductible_eur150
- sum_insured_eur60000
- claims_this_year1
Nothing yet. Use the widget and its events appear here.
These public demos run without an account and without AI calls. Switching widgets resets the demo.
Build by prompt, inspect and refine in the editor
- Edit HTML, CSS and JavaScript
- Define input schemas and data bindings
- Configure output events and action bindings
- Watch the result in a live sandbox with a bridge monitor
- Undo, redo and version history
Included
What every app gets without a single prompt
Governance
- Every change is a request with a risk score, reviewed by a human (human-in-the-loop, per app)
- Quality gate before approval: broken references, fields that don't exist
- Protected stages enforced in the database: one write path, with a reason and a before/after record
- Versioning: change history, numbered snapshots, export/import, one-click restore, disaster recovery
- Any number of stages, ten test stages if you like, and each can be protected. Staging compares and moves sections, hotfixes need a reason
- No server deployments: apps, stages and releases are configuration inside the platform
Users & access
- Onboarding from configuration: steps, roles, legal notices with consent
- Roles for customers and staff, sub-users and invite rights, teams
- Access per app and stage: open, invite-only or private
- Admin user management with masked personal data. Unmasking needs a reason and is logged
AI
- Agent routing: each question goes to a specialist agent with its own prompt, model and data rights
- Claude, GPT and Gemini, switched on or off per app and stage, with fallback
- Your own models through any OpenAI-compatible endpoint (Azure OpenAI, Mistral, self-hosted). The key stays on the server
- A model per platform function: the Architect, the router or the model council judge can each use a different one
- Masking before the model, emergency keywords before any AI, model council for sensitive questions
- AI calls of your users are counted, with daily limits for widget AI and invoice checks, and per-tier limits you can switch on
Automation & data
- Process engine (BPM) with a visual editor, human tasks and resume
- Triggers for schedules, data changes, webhooks and widget events, plus a task scheduler and a process monitor
- API connectors whose secrets never reach the browser
- Data kept apart per app and stage, export and erase on request, retention periods
One platform, several apps
For enterprise
Built for teams that need an audit trail
Atomar keeps a record of every proposed change and every decision about it. These parts are built and running today.
Available on request: SAML and other identity providers such as Okta or Google Workspace and apps on a domain of their own.
Limits
What the platform doesn't do
- The platform runs in the cloud. By default the server and the database sit in Frankfurt (EU), and the region is a setting. Masked text goes to the AI providers you switch on (Anthropic, OpenAI, Google). It is not local-first like the developer tools.
- Sign-in is with Google, and for companies with Microsoft Entra ID. SAML and other identity providers such as Okta or Google Workspace are available on request.
- Every app carries its own name, logo, colours, themes and wording. It runs under the platform address, and a domain of its own is available on request.
- Inputs on this page were frozen on the snapshot date. The results are computed by the platform's code when the page is built.
Under the hood, data is kept apart per app and stage, the database uses row-level security, and every API route that touches data checks who is asking. A test fails if a new route forgets to. Users can export or erase their data, and more than 7,600 automated tests run before every release.
Next step
See it for yourself
The example apps run on the same platform as this page. For a setup with your own sign-in, domain or models, write to us.