Boston’s CIO wants the public — and other city governments — to use his open-source agentic AI tools

Boston’s chief information officer, Santiago Garces, is betting on agentic artificial intelligence to transform the city’s open data into usable insights, and not just for data analysts, but for residents who are trying to better understand their communities.

To make that vision real, the city in October launched a model context protocol, or MCP, server — an Anthropic-built software layer that allows a Claude AI model to securely interact with data sources, like Boston’s open data portal. More recently, Garces built a set of open-source “skills,” Anthropic’s term for workflow templates, which run on the city’s MCP server. Garces said he thinks of each skill as a different AI “agent.” Each skill, or agent, sets up a complex analytical workflow that can produce formatted tables and plain language to answer various questions.

Developed with Bloomberg Philanthropies and the Abdul Latif Jameel Poverty Action Lab, or J-PAL, at the Massachusetts Institute of Technology, Garces said the tools are designed to let anyone access and analyze city data, with the goal of expanding transparency, improving policymaking and broadening who can participate in data-driven decision-making.

The MCP server, Garces said, acts as a foundational layer that allows the AI agents to interface with Boston’s open data system, which includes datasets on things like city infrastructure and public services. There are datasets on property records, trash schedules, snow emergency routes, public-school locations, tree canopy assessments, building permits and 311 service requests, such as pothole tracking.

Previously, a city data analyst may have had to download spreadsheets and then upload them into a chatbot to perform an analysis. But this process, Garces noted, was less secure, as it produced copies of data along the way, and it was less accurate because the data might be outdated — it’s just generally “more of a pain.”

Garces said it took some time to equip the MCP server to handle more complicated questions, but that it now gives AI agents a safe and structured way to query the city’s data. However, this function alone was not enough — the server needed context. Garces said that where the benefits of agentic AI became clear.

“The best people that do data analysis and policymaking kind of go through a certain workflow, or a set of workflows. So this is where terms like agentic AI come into play,” he said. “All an agent, or a set of agents, are is like a collection of things that work in a particular way, trying to solve a piece of the puzzle, if you will.”

These agents, he said, are guided with code written in plain language, step‑by‑step, on a number of tasks, such as how to frame a problem, decide what to compare, interrogate the numbers and turn them into clear findings so the system can answer complex questions instead of just spitting out raw statistics. They are informed via the habits of a good policy analyst, Garces added.

“All I did was take these really large bodies of work, like the Bloomberg Center for Public Innovation has like this handbook for public innovation. It’s like 50 pages, and it shows you how to train a person how to do this thing. I just condensed it to these files and Claude now follows that path,” he said. “These are the way that like really good data analysts and good policy analysts do things, we were trained to follow these kind of patterns of how to solve problems. It’s just that now I can push my AI tools to follow these recipes.”

One of the agents, which Garces called the “orchestrator,” reads the users’ questions and decides which of the other agents to activate and in what order. The orchestrator routes the question through the five agents — problem‑framing, analysis, communication, benchmarking or performance management — passing the output of each onto the next so the AI behaves more like a coordinated team of analysts, rather than a single chatbot.

“The goal here was, Can I level the playing field?” Garces said. “Can I have more people, both inside City Hall, and residents, be able to access kind of like that same level of thinking, if you will, like that same level of analysis that right now only like the mayor or like a department head would get access to, because you need access to a fairly sophisticated data or policy analyst.”

The skills, or agents, are available on Garces’ GitHub, but he plans to turn the technology into what he described as a shared, civic infrastructure. He said these agentic skills are just the beginning of what will be possible. In the coming months, Garces said, he plans to make Boston’s MCP server publicly accessible, so that anyone can use their own AI tools to draw on the city’s data to get answers to complex questions.

Garces is encouraging other cities to adapt the agents for their own open data portals. He said they’re compatible with the three largest open data platforms so that they can complete cross‑city comparisons. (The three most widely adopted open data platforms are developed by Tyler Technologies, Esri and Opendatasoft.)

In Boston, Garces said, his focus is getting every city employee who’s completed responsible‑AI training access to the agents, so frontline staff can do “mayor‑level” analysis without needing to wait on a data analyst.

“Our hope is then trying to figure out, how is it that we can get every city employee access to these tools? How is it that we can make sure that our constituents have access to these tools?” he said. “Because we just think that it can really help us have better conversations using the same data and the same facts as the ground truth, and it just, hopefully, helps us solve more complicated problems together in a way that is more civil, that is more interesting.”