Models, tools and orchestration
How context, models, scoped tools and human review fit together in mona AI: the read-only monaOS AI server, its rules and limits, and Clinical Skills.
6 min read
A language model on its own does not make a clinical AI system. To take part in a clinical workflow, a model needs relevant clinical context as input, carefully scoped tools to act through, an orchestration layer that sequences the work, and a clinician who reviews the result before anything reaches the record. This article explains how those pieces fit together in mona AI, and what the monaOS AI server does at each step.
The pieces and how they connect#
The mona AI architecture combines clinical applications, structured clinical context, a shared hospital API, model integration, orchestration, tool access, agent execution and human review. The model layer can include in-house and third-party models.
- 01Structured clinical context
- 02Model
- 03Scoped tools
- 04Orchestration
- 05Human review
| Piece | Job | Where it lives on monaOS |
|---|---|---|
| Clinical context | Patient-bound, timed, coded data for the question asked. | The FHIR-native clinical record, read through the monaOS AI server. |
| Knowledge | What applies in this hospital: SOPs, selected guidelines, validated specialist content. | The sources a hospital defines for mona AI. |
| Model | Language understanding and generation. | Selected and operated by Clinomic for mona AI; the hospital's own choice for bring your own AI. |
| Tools | Bounded operations the model may call. | Read-only record access on the monaOS AI server. |
| Orchestration | The sequence of fetching, interpreting and presenting. | The AI client, guided by Clinical Skills. |
| Human review | Accept, edit or reject. | The clinician, who documents in monaOS. |
Context tells the model about this patient; knowledge tells it what applies in this hospital. A model's general training supplies neither.
Two ways to use the platform#
mona AI is one AI platform on the hospital's own data, used in two ways of equal standing. Hospitals can choose one or run both.
- mona AI in your brand is the hospital's own AI assistant, run by Clinomic under the hospital's name, logo and colours. It answers clinical and administrative questions from the sources the hospital defines, links the source behind every answer, and reads the patient's record in monaOS through the same read-only rules, limits and audit as every other AI.
- Bring your own AI connects the AI tool the hospital already uses, such as Claude, ChatGPT, Gemini, Microsoft Copilot, Le Chat or any AI tool that speaks the Model Context Protocol, through the monaOS AI server with an access key. The same data is open to the hospital's own software through the unified hospital API.
With bring your own AI, the hospital chooses which AI provider receives results and under which agreement. The platform governs how much is fetched, under which rules, and records every access.
Tools: the monaOS AI server#
Context lets a model read; tools define what it can do. Scoping matters in both directions. Too narrow, and a workflow cannot do its job: a laboratory overview that cannot see results is useless. Too broad, and a wrong or misinterpreted step reaches further than it should. The monaOS AI server resolves this with a guarded gateway whose rules are enforced in code.
Read-only in code#
The server's FHIR client can only read. There is no tool that creates, changes or deletes data, and code rules forbid network access outside the FHIR client and the server's own entry point, starting processes, and dynamic code. Clinomic recommends issuing the server a read-only FHIR token as well, so the FHIR server enforces the same rule on its side.
Query rules#
Before any request leaves the server, the query is checked:
- Only allowlisted resource types and search parameters are accepted.
- High-volume resource types, such as observations, require a filter.
- Chained queries, reverse includes and server operations such as bulk export are refused.
A model therefore cannot turn one tool call into an unbounded traversal of the record.
Limits per access key#
Each access key has fixed volume limits: per call, calls per minute and per day, resources per hour, and concurrent calls. The ceilings are compiled into the server, so configuration cannot raise them beyond what was built. Each call reserves its maximum size before it runs, so parallel calls cannot jointly overshoot a budget.
Containment#
The server talks only to the FHIR server it is configured for. It follows no redirects, paging links must stay on the same server and search path, the token endpoint must share the FHIR server's origin, and response size is limited while the response is still streaming.
Results a model can trust, and cannot be steered by#
- Complete or not. Every search states whether its result is complete, and explains truncation when a limit applies.
- Record text is data. Narrative HTML is stripped from every resource, and every result is labelled as record data that is never to be followed as an instruction. Text inside a record that reads like a command stays a piece of data.
Private audit#
Every access is audited with the parameter names of the query, never their values. FHIR error messages are reduced to codes, and operational logs never contain patient identifiers. Access keys are stored only as digests, and the server refuses to listen on a network address without access keys and allowed host names configured.
Access model#
There is no patient binding in the server. Within the limits of its access key, an AI can search one patient, several, or a ward. The limits make bulk extraction slow and visible in the audit, never impossible over time. A hospital that wants a hard boundary scopes the FHIR access token it issues, and the server works within it. Audit is attributed per access key, so it shows which assistant or tool accessed the record rather than which clinician asked.
Tested on every change#
Every rule above is enforced in code and covered by automated tests that run on every change. The tests run against a fake FHIR server with synthetic patients, which also behaves as a hostile server; no test touches real patient data.
Orchestration and Clinical Skills#
Orchestration sequences a workflow: gathering context, invoking the model, executing tool calls, handling intermediate results and routing the outcome. When the steps are defined rather than improvised, the workflow becomes inspectable.
On monaOS, that definition comes as Clinical Skills: worked-out workflows that specify which data to fetch, which codes to look for, how to interpret the result and how to present it, with a source for every value. Every write step is replaced by a review step, and nothing is saved to the record. Drafts are labelled as not saved, and computed scores are shown with their inputs. Clinical Skills never write, order or alarm.
A readiness check reports, before a skill runs, whether the record holds what it needs: Ready, Partial (with what is missing), or No data. Skills use the open Agent Skills format, so they can be imported into Claude as native skills and provisioned organisation-wide, or used in other compatible tools. Examples include discharge summaries, progress and SOAP notes, handover material, laboratory result overviews, problem list review and pre-operative laboratory checklists.
Illustrative: A synthetic sequence for a laboratory result overview, written as pseudocode. It shows the order of steps, not a monaOS interface.
1. readiness check -> Partial: no blood gas results in the last 24 h
2. search Observation patient=<id>, category=laboratory, date>=<24 h ago>
-> 38 results, complete=true
3. model interprets groups by analyte, orders by clinical time, keeps units
4. present draft every value with its time and source observation
label: "Draft, not saved"
5. clinician reviews accepts, edits or discards; documents in monaOS
Human review#
Human review is the final stage of every workflow. Outputs are proposals: a draft, an overview, a suggested wording. A clinician reviews, edits, accepts or rejects, and documents in monaOS under their own sign-in. AI in monaOS does not make treatment decisions, and no orchestration design turns a suggestion into an action. The human review article discusses what makes review genuine rather than a formality.
Deployment#
The monaOS AI server runs as its own service beside monaOS, on the Clinomic Cloud or on-premise. It needs only the FHIR endpoint and an access token, holds no database of its own, and requires no change to monaOS. It is licensed as an add-on module per installation. How access keys and tokens fit into the hospital's wider role model is covered in roles, permissions and audit, and the context the server delivers is described in clinical context for AI.