open-fieldservice is an open-source field-service scheduler — a Preact front end, a Hono API, and a Cloudflare D1 database, billed as a self-hostable alternative to ServiceTitan or Jobber for pest control, HVAC, plumbing, and the rest. It ships with the full back office: jobs, a customer CRM, technicians, invoices, materials, a weekly calendar. What it didn't have was the one screen a dispatcher actually lives in all day — a board where you drag jobs onto technicians and watch the day fill up. So I cloned it, reskinned the whole thing ServiceTitan-style into "Atlas," and built that board. The scheduler, the API, and the data model are clawnify's open-source work — I want to be clear about that. What I added is the visual redesign and a new Dispatch board on top of the existing API.
What it is
open-fieldservice is Preact + TypeScript + Vite on the front end, a Hono REST API on the back, and Cloudflare D1 for storage (the project migrated to a D1-native pattern via a @clawnify/db dual-binding adapter, so it runs on Wrangler locally and deploys to Cloudflare). The data model is what you'd expect from a field-service app: customers, jobs, technicians, service types, invoices, materials, plus job notes and checklists. The API exposes the lot over about thirty REST endpoints, and there's a ?agent mode that swaps hover-to-reveal controls for always-visible buttons so an automation agent can drive it.
Atlas is two things layered on that base. The first is a reskin: a navy top app bar with a global search box, a navy sectioned sidebar (OPERATIONS up top, BUSINESS below), a Titan-blue (#2f6fed) accent, a KPI-card dashboard, and filled status pills instead of the original's bordered badges. The second, and the part I actually care about, is a new Dispatch board — technician rows crossed with a 7am-to-7pm timeline, where unassigned jobs sit in a left rail and you drag them onto a tech's track to assign and schedule them in one motion. Nothing about the database or the API changed; the board is pure front end reading the same endpoints the rest of the app already used.
How it was built
The reskin was deliberately surface-level. I didn't fork the architecture — I changed the shell (top bar plus a column layout), the palette and the component CSS, and the dashboard's stat cards into KPI cards. The point was to make it look like a commercial product without touching anything load-bearing, so the upstream code stays mergeable and the diff stays legible. It landed as an atlas-redesign branch merged to local main, then a second pass of nice-to-haves merged after.
The Dispatch board (src/client/components/dispatch-view.tsx, route /dispatch) is the real addition, and the constraint I set was that it could not require any backend work. It reuses the existing GET /api/schedule?start=D&end=D to load a day's jobs and PUT /api/jobs/:id to assign and reschedule — that's it. The board renders active technicians as rows and the 7am-7pm span as a timeline; each assigned job is an absolutely-positioned block whose left is computed from its scheduled_time and whose width comes from its duration. Unassigned jobs live in a rail on the left.
Assignment works two ways on purpose. The primary path is HTML5 drag-and-drop: drag a job from the rail (or an existing block) onto a tech's track, and the drop handler reads the cursor's horizontal offset, converts it to a time snapped to the nearest 15 minutes, and fires assignJob(jobId, techId, time) — one gesture sets both the technician and the schedule. The fallback is click-to-assign: click a job to select it, then click a tech row, which assigns without moving the time. I kept the click path specifically because it survives ?agent mode, where an automation agent can't reliably perform a drag.
Running it locally on Windows is its own small ritual, because this is a Cloudflare-stack app and I don't have pnpm installed globally. Two servers: Wrangler serves the API on 8787, Vite serves the UI (which lands on 5174 because 5173 was already taken), and D1 has to be seeded from schema.sql first. The npm dev script chains all of that with concurrently, but I ended up invoking the binaries directly — more on why below.
The gotchas
Three real ones, each of which cost time.
pnpm exec re-verifies the install and chokes, so I run the binaries directly. Installing deps with npx pnpm@latest install works, but trying to launch the servers through pnpm exec wrangler / pnpm exec vite made pnpm re-run its install verification, and it choked on native builds it had skipped. The fix is to skip the wrapper entirely and call the binaries by path: node node_modules/wrangler/bin/wrangler.js dev --port 8787 for the API and node node_modules/vite/bin/vite.js for the UI. Seed D1 the same way first: node node_modules/wrangler/bin/wrangler.js d1 execute open-fieldservice-db --local --file=src/server/schema.sql.
The Vite proxy resolved localhost to IPv6 and every API call 404'd. With the API on Wrangler and the UI on Vite, the dev proxy forwards /api/* to the back end. The proxy target was localhost:8787, but on this Windows box localhost resolves to IPv6 ::1 first, while Wrangler binds IPv4 127.0.0.1. So every /api/* request hit nothing and came back 404 — the app rendered but was completely data-blind. The fix was a one-liner in vite.config.ts: pin the proxy target to http://127.0.0.1:8787 instead of localhost. It's the kind of bug that looks like a broken API and is actually DNS.
"Today" is computed in UTC, so a behind-UTC clock lands on tomorrow. This is a pre-existing quirk that runs through every date-aware screen, the Dispatch board included: "today" is derived from new Date().toISOString(), which is UTC. On a clock that's behind UTC (most of the US for part of the day), that rolls the board and the dashboard onto tomorrow's date. I left it as-is rather than papering over it in one component, because the honest fix is a single shared local-date helper used everywhere, not a one-off patch in the board — but it's worth knowing about if you toggle to the dispatch view and wonder why it's empty.
What shipped
Atlas is merged to local main: the full ServiceTitan-style reskin (navy shell, sectioned sidebar, Titan-blue accent, KPI dashboard, filled pills) and the new Dispatch board with both drag-and-drop and click-to-assign, plus a global search dropdown in the top bar that filters across already-loaded jobs, customers, and invoices. None of it touched the schema or the API — the board is built entirely on the endpoints the upstream app already shipped, which is exactly what kept the change clean.
The division of credit, one more time: open-fieldservice — the Preact/Hono/D1 scheduler, the data model, and the thirty-odd REST endpoints — is clawnify's open-source project. My work is the Atlas reskin and the Dispatch board layered on top. A back-office scheduler now has the one screen a dispatcher actually wants to look at.
This is one post in a series on projects built this way. The running list is on the projects page.
