Contents
I stopped developing a custom data analysis agent and rebuilt the useful parts as an MCP server. Giving the team access to reliable data through the AI tools they already use was a more worthwhile investment.
Key Takeaways#
- Building a custom agent brings product responsibilities that extend far beyond analysis itself.
- An MCP server lets you provide specialized data and knowledge through existing general-purpose agents.
- Useful analysis starts before an agent receives a request, with agreed metric definitions and accurate data.
- Test workflows within the team’s existing tools before committing to building another application.
The Agent Became a Product#
The original goal was simple: ask questions in a chat interface, retrieve and interpret data, and visualize the results. I used Google Agent Development Kit (ADK) to build the agent and coordinate tasks.
But analysis was only part of the work. A practical application also needed a chat interface, authentication, session handling, persistent memory, file uploads, asset management, and rules for coordinating subagents. Developing each feature took time away from the core analysis capabilities.
Some of these could be skipped in a prototype. The issue was what users would compare it with. They already used Claude.ai, ChatGPT, and Gemini, so another application needed a reason to justify the effort of switching.
Memory Revealed the Costs#
Memory initially seemed straightforward: extract useful information from conversations, store it, and retrieve it when needed. The hard part was deciding what to remember and when past memories should inform new answers.
Corrections and deletions made the problem more complex. User preferences can change. Remembered assumptions can be wrong. Services like mem0 can reduce implementation work, but they do not eliminate the need to define and test this behavior.
There were also switching costs I could not resolve within my application. The context users had accumulated was already in the assistants they used. Rebuilding memory meant asking them to explain that context again from scratch, with no guarantee that the analytical benefits would justify the effort.
Putting Specialized Capabilities in MCP#
I replaced the entire application with an analysis server using Model Context Protocol (MCP). I narrowed its role to returning prepared data and explaining what it meant. Connecting it to a compatible general-purpose assistant provided capabilities specific to our work.
This changed the scope of development. The host application handles conversations and the available interface features. I focused on the data and expertise the assistant could not provide on its own. Visualization can also use host features such as Claude Artifacts where available.
If you make a similar move, keep the server’s initial scope small.
1. Expose one useful preaggregated dataset. 2. Provide definitions for its metrics and dimensions. 3. Explain its coverage, data freshness, and known limitations. 4. Connect it to an assistant your intended users already use.
Prepare the Data Before the Agent#
Initially, I thought connecting an AI model to BigQuery’s execute_sql tool might be enough to produce useful insights. It was not. Without enough preparation, the output tended to stick to obvious observations such as “trains are crowded during commuting hours.”
Meaningful progress came from work upstream of the model. We agreed on which metrics to track, prepared daily aggregates, and corrected how the source data was collected. These changes enabled the assistant to offer meaningful explanations. No amount of additional task coordination could replace them.
Before adding capabilities to an agent, check these four things.
- Has the team agreed on metric definitions?
- Is the necessary data being collected accurately?
- Are the aggregates reproducible and kept up to date?
- Can the assistant explain the limitations of the results?
Reuse Has Its Limits#
The transition did not require throwing everything away. Much of the analysis code remained useful, and I used FastMCP for the new server. Separating the data capabilities from the surrounding application also created opportunities to simplify the implementation.
After I shared the MCP server with the team through Claude.ai’s connectors, colleagues sent me examples of how they were using it. This was useful evidence of actual use, but it did not establish accuracy or measurable time savings. Those still need validation.
MCP does not remove every product responsibility. Data access still needs to respect permissions, and available features vary by host. There is a reason to build a custom application when existing hosts cannot support the workflows, interfaces, or controls you need. For our basic analysis needs, we found no reason to build an entire product.
The Next Step#
Choose one recurring analytical question and make it answerable through a single MCP tool in the assistant your team uses. Agree on the metric definitions and check the results against a trusted report.
Source: https://blog.mudatobunka.org/entry/2026/11/05/100000
Contents
