Large Language Models (LLMs) such as GPT-4 are rapidly entering classrooms as conversational tutors, yet their closed architectures leave four critical gaps for learning analytics: context is lost between sessions, interaction logs are unstructured, a single chatbot performs all functions without specialised traceability, and all data are stored on commercial servers outside institutional oversight. This paper explores how the open-standard Model Context Protocol (MCP) can bridge these gaps through a conceptual redesign of STEAMBrace Tester, an assistant that helps secondary-school teachers and educators refine STEAM activities. The proposed version replaces the cloud-hosted GPT with Claude Desktop running on institutional servers, links creator and evaluator agents through MCP, and records every exchange as xAPI statements in a local Learning Record Store. This configuration preserves each teacher's history, enables longitudinal analyses of activity quality, isolates the contribution of each agent, and ensures full data sovereignty under GDPR. Illustrative scenarios show how the enriched traces permit investigation of feedback uptake, the evolution of equity-oriented prompts, and the optimal number of iterative cycles. The resulting architecture offers a transferable pathway for educational institutions to reclaim analytic value from LLM-driven assistants while maintaining rigorous privacy and governance standards.
The emergence of large language models (LLMs), such as GPT-4, has popularized conversational assistants capable of generating and evaluating content of all types, including educational. These models offer insightful responses, but their "black box" architecture poses four critical barriers to
To illustrate this potential, this paper presents the STEAMBrace Tester case study, a customized GPT created in the STEAMBrace project (https://steambraceproject.eu/), funded by the European Horizon call, to help secondary school teachers and other educators improve STEAM activities. The current version, STEAMBrace Tester GPT, provides immediate value, but it is still imprisoned by previous barriers. This paper proposes the conceptual design of STEAMBrace Tester MCP, an evolution that leverages MCP to provide the tool with persistent memory, structured logging, and institutional governance of the data, thus expanding the spectrum of Learning Analytics available.
The objectives of this work are first, to analyze the limitations of LLMs with respect to LA and data sovereignty; secondly, to describe how MCP addresses these limitations; lastly, to demonstrate, using STEAMBrace, the pedagogical and analytical improvements obtained.
2.1.
The incorporation of large language models (LLMs) into teaching has led to conversational tutors,
material generators, and automated evaluators that provide near-instantaneous and differentiated
feedback. The Learning Analytics (LA) literature recognizes this potential, but stresses that the
analytics infrastructure remains immature and poorly integrated with educational research
processes [
In practice, LLMs face several obstacles that reduce their analytical usefulness. First, their
working memory is ephemeral: the context window-although expanding-continues to limit
continuity between sessions and hinders longitudinal tracking of learning [
These limitations are properly understood in light of the architecture of OpenAI's custom GPTs,
released in 2023. Each GPT is configured using three layers: (i) an instruction block-system prompt
that defines the desired behavior, (ii) a repository of knowledge documents that are attached as
additional context, and (iii) a set of tools or actions that empower the model to call external APIs
[
To address these limitations, Anthropic introduced in 2024 the Model Context Protocol (MCP),
described as a "USB-C for AI" because it standardizes bidirectional connections between models
and any external data source or tool, avoiding ad hoc integrations on a case-by-case basis [
The protocol is articulated in three logical components [
From an analytical perspective, MCP introduces four decisive capabilities. First, it enables
persistent context: before generating a response, the model queries profiles, histories or artifacts
stored outside its context window, overcoming ephemeral memory [
In order to understand the tool itself, first, it is necessary to explain the pedagogical context and objective of STEAMBrace (https://steambraceproject.eu/). The European STEAMBrace project aims to reduce the gaps in access and motivation towards STEAM careers in school citizenship. It articulates a hybrid methodology that combines Problem-Based Learning (PBL), Challenge-Based Learning (CBL), and Gagné's nine-phase principles methodology. It is also aligned with the principles of Design Thinking and Maker methodology for learning and prototyping orientation.
Within this framework, the project has developed the STEAMBrace Tester GPT tool, a conversational assistant designed for teachers and other educational agents teaching secondary school students who wish to share their existing STEAM activities and expand their proposals, identify biases (e.g. gender), make improvements, or adapt their proposals to curricular or extracurricular environments.
In the current architecture, the tool was built with OpenAI's Create a GPT function, which is based on three configurable layers: system instructions, knowledge documents and external tools. In the case of STEAMBrace GPT (see Figure 1), the system instructions define the evaluator role, the interaction process, and reference different shared files: the knowledge documents, which is composed of: the STEAM competency assessment rubric, templates of curricular and extracurricular activities and a rubric aligned with gender equity and inclusive language. Finally, the only tool enabled is the canvas mode, as the web search function has been omitted to eliminate hallucinations or errors in the processing of the system documents.
The workflow is as follows. In the Step 1, Language and Description, the teacher chooses the language in which to interact and receive feedback; and explains whether their activity is curricular or extracurricular, for the GPT to decide which template to use. In Step 2, Share, the teacher or educational agent shares the description of his/her activity (or creates it on the spot). In Step 3, Prompt -Chain, the GPT analyzes the text, checks it against the rubrics and assigns preliminary scores. In Step 4, Feedback, the GPT generates suggestions for improvement (variants, resources, evaluation indicators), and shares them with the user. Lastly, in Step 5, Fine-Tunning, the teacher iterates until a satisfactory version is obtained and may request a resource pack (e.g., link to Canva, Maker materials list).
Despite its initial adoption planned for Q2-2025, the prototype exhibits the same constraints that affect commercial LLMs, as previously discussed. The assistant loses its internal memory upon logout, meaning it does not retain previously reviewed activities, which complicates any effort to track a teacher’s progress over time. Additionally, all interaction logs—including messages, timestamps, and revisions—are stored as unstructured plain text on OpenAI servers. Analyzing the sequence of interactions either at the individual or collective level would require manual extraction and structuring, which is incompatible with xAPI standards. Furthermore, the model operates as a single agent: it generates and evaluates content but lacks the capacity to delegate tasks to specialized agents or to share task status across models. This restricts the potential for collaborative evaluation processes involving agents focused on specific areas such as PBL, equity, or particular STEM disciplines. Finally, data governance remains in the hands of the provider. Project administrators only receive summary reports, while full interaction traceability is not accessible to the research team, thus hindering the ability to conduct rigorous impact assessments. For instance, although the tool is currently being used by over 50 educators across Europe, aggregate-level information on user interactions remains unavailable.
These limitations motivate the design of the STEAMBrace Tester MCP version, described in the following section.
The STEAMBrace Tester MCP version represents a conceptual evolution of the educational assistant, replacing the cloud-hosted GPT with a Claude Desktop model deployed on in-house servers, connected via the Model Context Protocol (MCP) to internal services managed by the institution. While the user experience remains that of an accessible and friendly conversational assistant, operations that previously occurred opaquely within the commercial provider's infrastructure are now distributed in independent components that favor personalization, data control and analytical exploitation from the Learning Analytics framework.
The following is a summary (see Table 1) of the main phases of the STEAMBrace Tester MCP workflow, together with the analytical functionalities that are activated in each of them:
From the first step of the interaction, specific functionalities are activated. When the teacher selects the language and specifies whether the activity is curricular or extracurricular, the system accesses a local memory base where it consults previous activities linked to the same user. This persistent memory allows the model to adapt its recommendations according to the history, generating pedagogical continuity and longitudinal data that can be analyzed later to study, for example, the evolution of quality criteria or the progressive incorporation of equity elements.
In the second step, when the teacher shares his activity (either written from scratch or adapted), the main analysis flow is activated. From that moment on, each exchange of messages, settings or questions is automatically recorded in a Learning Record Store (LRS) using statements structured under the xAPI standard. This includes not only the text of the interactions, but also metadata such as the response time, the type of suggestion requested or the latency between actions. This traceability makes the wizard an instrumented tool, capable of providing evidence at both the individual and institutional level.
In the following steps, when the model generates feedback and suggestions, this content is linked to intermediate versions of the activity, which are also stored and tagged in the LRS. In this way, not only the final version of the proposal is preserved, but also the entire iterative process that produced it. This information makes it possible to study the number of improvement cycles, the evolution of quality according to the STEAMBrace rubric, or even the sequence of actions that generate the most added value.
Finally, in the final review phase, the system continues to record any future interaction. Moreover, having functionally separated the agents, it is possible to refer the final proposal to a different evaluation model (e.g., a Llama-3 refined to apply the rubric or detect gender bias), allowing a specialized and complementary evaluation to the creative assistant. Both interventions— creation and evaluation—are recorded separately, facilitating their comparative analysis.
This architecture also allows non-invasive collection of complementary traces: clicks on suggested external resources, permanence at each step, language changes, use of templates... All these variables are securely stored in an on-premises environment, encrypted and under institutional control, which facilitates GDPR compliance and reinforces the sovereignty of educational data.
The architecture presented in the previous section allows for a qualitative leap in the analysis and understanding of teaching work in the specific context of STEAM. Beyond solving the technical limitations already identified, STEAMBrace Tester MCP opens the possibility of developing new lines of analysis and functions that were previously unattainable.
For example, with the data collected by the LRS it is possible to reconstruct the complete cycle of design, revision and improvement of a STEAM activity. This makes it possible to analyze how many iterations a teacher performs, what kind of suggestions he/she accepts and how his/her score on the rubric evolves after each change. With this traceability, customized mentoring programs could be designed based on real improvement trajectories, or identify critical moments where creativity or the integration of approaches such as CBL stagnates.
Another possibility lies in the aggregate analysis by type of teacher or center. The database could be linked to variables such as educational level, teaching seniority, or rural/urban context, making it possible to study whether certain profiles respond differently to certain suggestions, or whether the intensive use of the assistant correlates with a greater diversity of activities or with a sustained improvement in equity criteria.
From a pedagogical point of view, new forms of support could be activated. For example, if a teacher is developing an activity with a Maker approach but shows doubts or setbacks in the versions, the system could suggest specific support resources or connect him/her with another user who has overcome similar difficulties. These recommendations could be managed by specialized agents, integrated through MCP, without the user perceiving a greater complexity.
In terms of educational research, the repository of structured and contextualized data would make it possible to advance in questions that have not yet been explored, such as the impact of AI assistants on teacher self-regulation, or the way in which different design sequences -for example, generation first, evaluation later, or vice versa- affect the final quality of the activities. In addition, the use of inclusive language, the representation of female referents in the activities or the explicit attention to equity issues could be studied, extracting patterns from the processed texts themselves.
Although there are currently initiatives such as Playlab.ai [
Therefore, in addition to being a technical solution to previous barriers, STEAMBrace Tester MCP is configured as a living analytical infrastructure, capable of generating applicable educational knowledge and facilitating informed pedagogical decisions. This shift from a closed and opaque architecture to an extensible, governed and learning-oriented system represents a relevant contribution to the debate on the responsible use of AI in education.
Acknowledgments
Funded by the European Union—European Innovation Council
STEAMBrace project—Grant Agreement nr. 101132652. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Innovation Council. Neither the European Union nor the granting authority can be held responsible for them.
During the preparation of this work, the authors used DeepL and ChatGPT-4 in order to: grammar, spelling and overall improvements of translation. After using these tools/services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [20] Poker-Edge.Com, Stats and analysis, 2006. URL: http://www.pkredge.com/statsYYFWWQ.php. [21] B. Obama, A more perfect union, Video, 2008. URL: http://video.google.com/videoplay? docid=6528042696351994555. [22] D. Novak, Solder man, in: ACM SIGGRAPH 2003 Video Review on Animation theater Program: Part I - Vol. 145 (July 27–27, 2003), ACM Press, New York, NY, 2003, p. 4. URL: http://video.google.com/videoplay?docid=6528042696351994555. doi:99.9999/woot07-S422. [23] N. Lee, Interview with bill kinder: January 13, 2005, Comput. Entertain. 3 (2005).
doi:10.1145/1057270.1057278. [24] J. Scientist, The fountain of youth, 2009. Patent No. 12345, Filed July 1st., 2008, Issued Aug.
9th., 2009. [25] B. Rous, The enabling of digital libraries, Digital Libraries 12 (2008). To appear. [26] M. Saeedi, M. S. Zamani, M. Sedighi, A library-based synthesis methodology for reversible logic, Microelectron. J. 41 (2010) 185–194. [27] M. Saeedi, M. S. Zamani, M. Sedighi, Z. Sasanian, Synthesis of reversible circuit using cyclebased approach, J. Emerg. Technol. Comput. Syst. 6 (2010). [28] M. Kirschmer, J. Voight, Algorithmic enumeration of ideal classes for quaternion orders, SIAM J. Comput. 39 (2010) 1714–1747. URL: http://dx.doi.org/10.1137/080734467. doi:10.1137/080734467. [29] L. Hörmander, The analysis of linear partial differential operators. IV, volume 275 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], Springer-Verlag, Berlin, Germany, 1985. Fourier integral operators. [30] L. Hörmander, The analysis of linear partial differential operators. III, volume 275 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], Springer-Verlag, Berlin, Germany, 1985. Pseudodierential operators. [31] IEEE, Ieee tcsc executive committee, in: Proceedings of the IEEE International Conference on Web Services, ICWS ’04, IEEE Computer Society, Washington, DC, USA, 2004, pp. 21–22. doi:10.1109/ICWS.2004.64. [32] TUG, Institutional members of the TEX users group, 2017. URL: http://www.tug.org/instmem.html . [33] R Core Team, R: A language and environment for statistical computing, 2019. URL: https://www.R-project.org. [34] S. Anzaroot, A. McCallum, UMass citation field extraction dataset, 2013. URL: http://www.iesl.cs.umass.edu/data/data-umasscitationfield.