When we shipped Healthie Dev Assist last year, the goal was simple: bridge the gap between natural language and Healthie's GraphQL API. Ask Claude or ChatGPT a question about the schema, get back precise, executable code. No more sifting through docs trying to figure out what fields createAppointment actually accepts.

It worked. But as usage grew, and as AI tooling costs stopped being rounding errors, the original architecture started showing its ceiling.

Every schema exploration task required the AI tool to make sequential round-trips: search for a type, wait. Introspect a field, wait. Follow a reference, wait again. For complex cross-type queries, this compounded quickly. More turns meant more latency, more tokens consumed, and more noise in the UI from intermediate tool outputs.

Dev Assist 2.0 fixes this at the architectural level. It's not a patch, it's a rebuild of the execution model. And it ships today as a drop-in replacement for anyone already running 1.0.

‍

📌 Dev Assist is the AI layer that lets your engineers talk to Healthie's API in plain language. 2.0 makes it significantly faster and cheaper to run, without any changes to how your team uses it.

A Quick Recap: What Dev Assist Is (and How MCP Fits In)

Healthie's GraphQL API is comprehensive by design. Scheduling, charting, patient management, billing, care plans, it's a deep surface area, intentionally built for teams that want to compose sophisticated care delivery workflows rather than accept off-the-shelf constraints.

That depth comes with a tradeoff: there's a lot to know. Even experienced developers spend meaningful time exploring the schema, figuring out which types connect to which, how to construct the right mutation, or what the response shape will look like.

Dev Assist removes that friction. It connects AI tools like Claude, Cursor, and OpenAI directly to Healthie's API via the Model Context Protocol (MCP).

‍

🔌  MCP (Model Context Protocol) is an open standard that allows AI assistants to interface with external tools, APIs, and data sources in a structured way, rather than through copy-pasted documentation. Think of it as a live, queryable bridge between an LLM and a software system.

‍

With Dev Assist running, you can ask your AI tool things like:

"What fields does createAppointment accept?"

"How do I query a patient's insurance coverage?"

"Show me all billing-related types across the API."

‍

And receive precise, tested, executable code, not a hallucinated guess based on documentation that may be months out of date. Dev Assist works from a locally cached copy of Healthie's GraphQL schema, so the types, fields, and relationships it reasons over reflect the actual API.

This is the foundation. Dev Assist 2.0 is what happens when you go back and rebuild that foundation to be significantly more efficient.

Introducing Dev Assist 2.0: A New Execution Model

The original Dev Assist was built around two tools: one to search the schema, one to look up type details. The catch? Claude or ChatGPT could only use them one at a time, search, wait, look up, wait, follow a reference, wait again. Every question required multiple back-and-forth exchanges before you got an answer.

Dev Assist 2.0 changes how this works under the hood. Instead of sequential steps, it now runs all schema operations in a single parallel block, so Claude or ChatGPT can explore multiple types at once and return a complete answer in one shot.

The result:

• 64% fewer tokens consumed per session
• 55% fewer round-trips
• ~5x faster on average, up to 84% faster on complex queries involving billing and insurance types

What else is new:

• Live API execution — Dev Assist can now run actual test queries against the Healthie API and return real response shapes, not just what the schema says a field should accept. This closes the loop between exploring the API and seeing what it actually does.
• Cleaner responses — answers come back as structured data instead of raw schema text, which means more accurate, easier-to-use results.
• Cross-platform — removed a macOS-only dependency. Dev Assist now runs on any OS.

‍

💰  The practical translation is that your engineers get answers faster, the AI makes fewer mistakes on complex multi-type questions, and, if you're running on a paid Claude or ChatGPT API plan, it costs meaningfully less per session.

What Changed Under the Hood

Several architectural decisions in 2.0 are worth understanding if you maintain or configure the server:

• Binary dependency removed. The original server wrapped apollo-mcp-server, a 15MB binary, macOS-only, invoked via child_process.spawn. Dev Assist 2.0 is pure Node.js + TypeScript. No subprocess. No platform constraint. It runs on any OS.

• Single tool instead of two. Consolidating search_schema + introspect into execute_healthie_code removes the overhead of Claude or ChatGPT selecting between tools and eliminates the constraint that each operation must be a separate LLM turn.

• Structured JSON, not raw SDL text. The old introspect returned SDL fragments, text that Claude or ChatGPT had to parse to understand field names, types, and relationships. The new version returns structured JSON objects, which models reason over significantly better.

• Pagination on large types. The User type has 455 fields. Without limits, introspecting it returns ~40K tokens. The new limit option returns the first N fields plus a totalFields count, letting Claude or ChatGPT self-regulate context consumption on deep types.

• Live API execution. Dev Assist 2.0 adds healthie.query() and healthie.mutate(), allowing Claude or ChatGPT to run actual queries against the Healthie API and return real response shapes, not just inferred schema definitions. This unlocks a class of questions that were previously unanswerable: not just what does this field accept, but what does this call actually return.

Environment variable isolation. After loading configuration, all environment variables are cleared from the process. API keys are loaded once and then removed from the environment, limiting exposure if sandbox code escapes the vm context.

Getting Started

If you're upgrading from Dev Assist 1.0

This is a true drop-in replacement. Your existing configuration works without modification:

1. Pull the updated server code from GitHub
2. Restart the MCP server (or Claude Desktop)

If you're getting started with Dev Assist for the first time

Dev Assist connects to Healthie's sandbox environment. You'll need basic familiarity with git and npm. Setup takes under five minutes.

3. Visit github.com/healthie/healthie-dev-assist
4. Follow the Quick Start Guide in the README
5. Connect your preferred AI tool (Claude, Cursor, OpenAI) via MCP
6. Start exploring the schema in plain language

For full API documentation and production implementation details, visit docs.gethealthie.com.

The Bigger Picture

Dev Assist 2.0 is part of Intelligence by Healthie, our ongoing investment in making Healthie's infrastructure AI-native, not AI-adjacent. The same architectural choices that enable parallel schema traversal today are what position the platform to work with increasingly capable AI agents as they emerge.

Healthcare infrastructure that requires sequential, one-step-at-a-time interaction with AI tooling isn't going to scale with where the technology is heading. The platforms that will compound over the next several years are the ones where AI can operate fluidly, reading context, following relationships, executing complex workflows, without hitting architectural ceilings.

Dev Assist 2.0 removes one of those ceilings. We're just getting started.