Yersinia pestis can infect the lungs, causing pneumonic plague. The disease can be transmitted between humans through infected respiratory droplets, and symptoms can appear in a very short incubation period, says the WHO. Pneumonic plague can cause death within 18 to 24 hours of disease onset if untreated. Antibiotic therapy can cure the infection when administered early. WHO also notes that verification includes laboratory tests including identification of Y. pestis from specimens such as blood or sputum.
That leaves a specific role for computing. Software can collect reports, find related records, verify if several articles come from the same source, compare dates and retrieve the latest laboratory status. It cannot substitute for a test of whether Y. pestis is present. The proposed workflow separates those responsibilities, so that a machine-generated summary cannot silently turn a suspicion into a diagnosis.
Architecture of MCP and AI workflow proposal
Disease report or clinical signal
↓
Public health records, laboratory systems, hospitals, genomics, GIS and One Health records
↓
Separate MCP servers
↓
Identity checks, access permissions and request logging
↓
AI evidence processor
↓
Source tracing, duplicate detection, entity matching, contradiction checks and location matching
↓
Reported → Investigating → Laboratory Pending → Confirmed / Ruled Out
↓
Epidemiologist, microbiologist and clinical review
↓
Approved medical and public-health response
In the design, the MCP is applied as a connection method between the AI application and systems that already have specialized information. MCP servers can expose resources that provide reference data and tools that perform defined actions or queries. In the current MCP documentation, resources are contextual data provided to clients, and tools are callable functions that can retrieve information or perform operations. OAuth-based authorization can also be used to implement access control on the whole server or on specific tools.
The AI application wouldn’t have free access to every hospital record, laboratory database or genomic file. Each MCP Server may expose only the fields required for the task it is assigned. For example, a laboratory service may return test status without disclosing the rest of a patient’s medical record. This keeps the language model on the information side of the process while clinical and laboratory systems keep control of medical evidence.
Step 1: Get the first report
The monitoring service could be based on data from health authorities, hospitals, laboratories, veterinary services and selected public sources. Text parsing could identify references to pneumonic plague, Yersinia pestis, severe pneumonia, deaths of animals, laboratory testing and locations associated with an event. The system would generate a record of the investigation with the source and time of publication attached. The mention of “pneumonic plague” in the source would not necessarily make it a confirmed case of plague.
Step 2: Follow the original source
Repeated claims can appear more credible online than they are. News aggregators, discussion sites, social accounts and videos can all copy one article, creating hundreds of pages based on the same original statement. A source-tracing service could gather those copies and hold the earliest identifiable origin of the claim. And the AI would realize that 300 repetitions of one statement is still one evidence chain.
Step 3: Query the laboratory system
A Laboratory MCP Server may be able to provide a limited query for the status of a relevant specimen. The answer can be received, testing, not detected, presumptive or confirmed. The terminology is those selected by the laboratory itself. The AI would read that status and add it to the investigation record. It would not generate its own lab result from symptoms or newspaper coverage.
Plague guidance from WHO says laboratory testing is the point of confirmation. Its fact sheet on September 2026 says laboratory testing is needed to confirm plague and that Y. pestis can be found in pus, blood or sputum. WHO’s 2024 surveillance manual also includes procedures for the diagnosis, treatment and surveillance of plague. An authorized result would be transported from the laboratory to the application reviewing the incident only via an MCP connection.
Step 4: Use genomics post-pathogen identification
If Yersinia pestis is confirmed, sequence analysis may be relevant for the investigation. Laboratory sequencing and bioinformatics can compare isolates and give specialists information to study whether samples might be related. The AI does not have to calculate those relationships from raw genome files. A genomic MCP server could produce an approved analytical result by software and reviewed by the laboratory’s own procedures.
Such a distinction prevents treating the general language model as a genomic-analysis software. The specialist service calculates this scientifically, stores the source material and returns a defined result. The artificial intelligence can place that result next to the clinical, geographical and laboratory records. Then epidemiologists can consider how much weight they want to give it in their evaluation of the case.
Step 5: Check animal watch and geography
Animals and fleas are also involved in plague surveillance as Yersinia pestis is a zoonotic bacterium. According to the WHO, plague usually occurs among small mammals and their fleas, and humans are infected by flea bites, contact with infected materials or by inhaling respiratory droplets of a patient with pneumonic plague. WHO plague guidance recommends tracing contacts and investigating probable sources of infection. So the information from human medicine alone may not be enough for an epidemiological investigation.
A GIS service may be able to determine if reported human cases, animal findings or exposure locations fall within a defined area and period. Permitted veterinary or vector records could be provided by a separate One Health service. The AI could then cluster those records for examination without implying that two events close by had to be related to each other. The location match provides investigators with a question, but not proof of transmission.
6. Find conflicts between reports
Disease investigations change as new information comes in. A first report may indicate suspected pneumonic plague, a lab record may indicate that Y. pestis was not detected, and a health authority may update the incident status. An evidence processor can compare newer records to earlier claims, and can mark which statement is still current. The previous report can remain in the history and not appear again as the actual medical status.
A case record could be reported, investigating, laboratory pending, and then confirmed or ruled out. Each update will show the source that caused the status update. Instead of depending on an old article in its previous conversation context, the AI would ask for the current state whenever it is preparing a new report. This makes it easier to track corrections when a fast moving health story changes.
Step 7. Present evidence to competent reviewers
The workflow completes its machine-processing stage before medical action is approved. “Epidemiologists look at transmission, microbiologists look at pathogen testing and clinicians decide how a patient should be managed.” The artificial intelligence can put together a short record of what is confirmed, what is still being tested, where sources disagree and what records are missing. It should not be able to independently declare a plague outbreak or determine treatment for a patient.
WHO guidance for pneumonic plague includes patient isolation, contact monitoring, specimen collection and early antibiotic treatment. But those are a matter of medical and public-health judgment, not a language model executing a workflow. The software can monitor whether or not an authorized action has been completed and bring unresolved tasks to the attention of the relevant team. Clinical authority is not in the picture.
What MCP offers
MCP provides a consistent way for an AI application to access resources and invoke allowed tools outside the language model. A public-health service might give an incident status, a lab service a test result, and a GIS service can respond to a location query. Authorization can limit access to particular servers or specific tools and this is important when health information is concerned. MCP does not diagnose pneumonic plague, test biological samples, or decide what public health response should be.
The design proposed assigns each job to the system that is qualified to perform it. Labs identify the pathogen, genomic software analyzes the sequence, GIS software performs the spatial calculations, and health professionals interpret the combined record. Artificial intelligence reads across those outputs, prepares the information for review. MCP defines the way those requests and responses can flow under defined permissions.
Artificial intelligence in plague surveillance: a practical approach
The useful target is not an AI model that guesses from internet activity whether a patient has pneumonic plague. A better target is software that can find the original claim, check the latest authorized record, notice when two sources disagree, retrieve the laboratory state, and show investigators what changed. That task is technical, measurable and consistent with the way medical authority already operates. It also makes it less likely that repeated online phrasing turns into an unsupported outbreak statement.