THE THING WE BUILT:

Chief

When you run an agency, the hard part isn’t having data (it’s everywhere.)

The hard part is chasing it down (across Harvest, Xero, Notion, Linear — take your pick).

If I’m honest, I’m both the culprit of our data’s messiness and the main person responsible for sifting through it.

As a software engineer, I’ve put together an AI-enabled tech stack that allows me to send a message that sounds like,

“Hey Chief… can you just quickly run a one-, three-, and ten-year financial projection for us?”

You know — the kind of “quick” request that somehow eats an afternoon and at least one teammate’s patience.

So we built a thing.

Chief in action

How It Works

Chief is an AI-powered assistant that lets me delegate those questions to a cluster of agents instead of people.

By connecting a series of agents to our accounting, time-tracking, and project systems, we can ask questions in plain English…and get answers in plain English. Text-based answers are supported by visualizations when it matters. Plus, there’s always the option to save key insights straight to your dashboard.

1. Ask in plain English.

   
   

   “What did we invoice Client X for in Q3?” Chief queries its data sources (Xero, Harvest, Notion, etc) through the Model Context Protocol (MCP).

   
   

2. Chief answers in English.

   
   

3. Pair it with an appropriate visualization.

   
   

   If Chief decides there’s interesting data to be shown, it automatically creates a visualization. Based on the data it’s reviewing, it chooses and formats the best chart (line for trends, bar for comparisons, area for cumulative totals)

   
   

4. Save relevant insights.

   
   

   Every chart or response can be pinned to your dashboard with one “Add to Dashboard” click.

   
   

5. Keep it live.

   
   

   Chief remembers the original prompts used to generate each dashboard component and can refresh the data on demand. This keeps metrics up-to-date without re-querying everything from scratch. That means faster updates, lower token usage, and dashboards that stay relevant.

   
   

6. Work in context.

   
   

   Chief’s split-screen interface lets you chat and view your dashboard simultaneously. Discover, visualize, and keep learning — all in one place.

   
   

Notes from Building

1. Discovery Should Drive Design.

   
   

   Most dashboards start with assumptions about the metrics we think matter. Chief inverts that. By turning every question into a candidate for visualization, we built dashboards that actually reflect what people care about. The “blank dashboard problem” from traditional Business Intelligence disappears when insight precedes configuration.

   
   

2. Frameworks Aren’t Equal.

   
   

   In building Chief, we evaluated both the Claude Agent SDK and the OpenAI Agents SDK. Our experience echoed what we saw in industry comparisons: the Claude SDK offered smoother agent-to-sub-agent orchestration, stronger built-in context management, and a more cohesive agentic workflow overall.

   
   

   Claude’s models (Opus, Sonnet, and Haiku) are deeply aligned with the SDK’s native tool-use patterns, making loops like reasoning → tool → reflection feel seamless. The framework also handled context automatically, compacting older messages without losing important information. With the OpenAI SDK, we had to build that logic ourselves — manually managing history truncation and token budgets.

   
   

   While GPT-5’s reasoning was undeniably strong, the overall agentic experience with Claude felt more production-ready. In the end, the choice came down to focus: less time debugging orchestration, more time building the product.

   
   

3. MCP Changes the Game.

   
   

   Model Context Protocol made data integration almost frictionless. Each data source (Xero, Harvest, Notion) runs as a self-contained “context server,” so adding a new integration is as simple as wiring up a new endpoint. It’s modular, predictable, and future-proof.

   
   

   Technically, you could recreate parts of this setup yourself by wiring MCP servers directly into Claude or ChatGPT. But Chief adds a few layers that make it actually usable in practice like:

   
   

   • A simplified MCP configuration flow

   • Context-aware agent instructions that improve how each service is queried

   • The conversational dashboard that ties everything together

   
   

   The result feels less like a proof-of-concept and more like a real workspace.

   
   

4. Interfaces Should Feel Human.

   
   

   We designed Chief to stream responses, show real-time tool calls (“Xero: Get Invoices”), and render charts inline. Those small touches (transparency, immediacy, clarity) build trust. When people can see how the AI is thinking, they engage more deeply with the answers.

   
   

Chief was built to save time.Instead, it made us realize how much more there was to ask.

It’s a bit of Jevons’ Paradox in action — the easier it became to get answers, the more questions we found ourselves asking.

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.