Anvil2 Design System
- Company
- ServiceTitan
- Year
- 2022–present
- Role
- [your role]
Intro
From a system team to an infrastructure platform
I like to think of this case study as the evolution from a system team to an infrastructure platform team: how we built a foundation that could support a high-growth environment like ServiceTitan, make a direct impact on users, and support a flood of product and feature launches.
The problem
Anvil1
Anvil1 was the system ServiceTitan had when I joined. Its core problem was architectural: any prop could pass through, so nothing stayed consistent, components got overcomplicated, and users complained — the clearest one, having to tab into a dropdown just to search it. Eighty percent adoption wasn’t proof the system worked; it was masking inaccessibility and only about 40% reliable responsive behavior.
Building Anvil2
Building Anvil2
We couldn’t create the guardrails we wanted inside Anvil1 without disrupting people. When you give that much flexibility, it’s hard to take back. A small, lean team shipped v1 in about eight months, pre-AI, starting from the atomic elements: more prescriptive, but structurally flexible. Distinct token sets apply across modes, themes, illustrations, components, and kits — domain-specific packages that customize while protecting the core — including white labeling and the evolving agentic experience at ServiceTitan.
Craft
A real investment in craft
The system also got real investment in craft: enhanced color ramps, a standardized illustration language (heroes and empty states), data visualization guidance, clear iconography, and contribution paths powered by AI. Bumping our color ramp from 650 to 900 opened up customization, white labeling, and more.
Partnerships
Engineering was the hardest hill to climb
None of that happened in a vacuum. Product, design, and engineering had to commit to upgrades; engineering was actually the hardest hill to climb. We brought engineering into beta testing, partnered with the platform team on versions inside the monolith to increase stability and reduce bugs substantially, and built a pipeline for breaking changes and releases. We also pushed for an ADA roadmap, overhauled visual design, brought in motion, and partnered with content on voice and tone.
Adoption & impact
Buy-in is one thing. Use is another.
Getting buy-in is one part of it; day-to-day use is another. We invested in migration tools, codemods, weekly office hours, and design technologists embedded on our highest-impact products. Since launch, it’s supported app launches, personalization, and internationalization, and it’s the system ServiceTitan uses to power agentic and GenUI experiences.
When we measured teams with 6 developers and 2 designers, teams saw an average of 192% ROI. The developers estimated they saved about 33% of their time. The designers said our resources doubled their speed.
Wired for AI
Wired for AI
One of our first moves was switching to a docs platform sourced straight to markdown, so an agent could parse and query the system — a good introduction to MCPs, and a way to support human-to-AI collaboration. Our support channel quieted down, prototypes got more aligned, and adoption increased. We also learned we could provide better content and structure than Mintlify’s MDX allowed.
Powered by AI
Powered by AI
So we built our own orchestrator, beyond docs. Treating AI as a first-class consumer, it uses multiple tool calls from sources we feed the MCP: tests, migration data, PRs to help teams migrate, accessibility fixes, token-pipeline consistency, better MDX, and more accurate prototypes. We’re still figuring a lot of this out, honestly. The clearest lesson: the team has extended past a traditional design system into infrastructure, producing skills that power agentic bots and support R&D and product across ServiceTitan.
Principles
Principles
These are the six principles we hold ourselves to. They’re documented, and we check new decisions against them.
- Consistency that scales
- Efficient and sustainable
- Built for everyone
- Quality you can trust
- Provide and curate
- Anchor the core, enable the rest
Coda: exploring GenUI
Exploring GenUI
I got the itch to explore GenUI myself. I built a short booking-flow demo with an AG-UI bridge that relies solely on Anvil guidance to render the right component at the right moment, using a base selection of components and the docs MCP server. It became the catalyst for GenUI in ServiceTitan’s Atlas product.
[TODO] The script says “March,” with no year stated.
- The agent reads the conversation and decides which tool to call first — no hardcoded step order.
- It queries Anvil2 and picked a dialog, because the user needed to take an action, not browse (as they might in a drawer).
- That selection triggers an AG-UI event, and the agent uses a dialog again for a blocking confirmation.
- I built this in March, and we’ve come leaps and bounds since.