RouteListToTesla turns ordered address screenshots into reviewed Tesla routes. The browser uploads selected images to the application server, where Tesseract, PaddleOCR, and EasyOCR independently read them and produce a majority transcription with literal alternatives for review. Google Maps geocoding is a separate service.
The container builds from source in separate Java and OCR stages. Production uses
a stripped Java 25 jlink runtime as non-root UID/GID 10001, with no JDK or Maven
in the final image. Build and validation commands are in the
development reference.
- In TAP, open Other Tools → TeslaRouter, or open TeslaRouter directly, then tap Continue with TAP.
- Sign in with TAP and approve RouteList access in TAP. TAP controls identity, billing, the
routelistentitlement, and per-subject vehicle grants. - In Import your stops, select an authorized Vehicle. Set Default state only when a screenshot omits its state.
- Tap Choose route screenshots. On iPhone, choose Photo Library and select multiple screenshots. Reorder them in the selected-file list before tapping Read addresses; their order determines route-stop order.
- In Review & send, check the OCR candidates, correct text, and remove unwanted addresses. Confirm or edit flagged readings before sending. The page does not support creating a new address by typing or importing a text list.
- Send a route group with Send Route 1 to Tesla; the number changes for each group. Tesla routes are grouped to respect the eight-waypoint limit. Automatic navigation is optional: select Use Automatic Navigation, then Start Automatic Navigation.
TAP grants last up to eight hours, have no refresh token, and are held in memory. Expiry or an application
restart requires signing in with TAP again. TAP checks current entitlements and vehicle grants; revoked access
blocks browser and navigation requests. RouteList uses the stable TAP subject (sub) as identity. TAP email is
unverified profile data and is not an authorization key.
TeslaRouter shares TAP's design and links back to TAP Dashboard and Other Tools. On narrow screens, open Menu to find these links. See the design and deployment record.
Use Safari and its native Photos picker as the primary flow; the site does not read the Photos library in the background. For an iOS 26+ Home Screen shortcut that opens in the browser, leave Open as Web App off. Standalone mode remains optional and unverified on a physical iPhone. See iPhone setup guide for the steps. Shortcuts, share-sheet import, and a native iPhone app remain proposals and are not implemented.
Image cache checks are owner-scoped, use SHA-256, and preserve selected-image ordering. All three OCR
engines run concurrently and locally with models installed at build time; no OpenAI key or hosted OCR is used. Review and editing remain
user-controlled. Overlapping screenshots merge matching consecutive address rows, preserving distinct
units and repeats within a screenshot. Google Maps geocoding stays separately configured with the existing google.api.key setting.
Use Java 25 and the Maven wrapper. Real OCR requires the three-engine wrapper and its offline models; the Dockerfile installs them. Plain Tesseract TSV is accepted only by explicit legacy tests. Configure TAP delegated login and existing geocoding settings as described in TAP setup and the development reference .
./mvnw spring-boot:run
./mvnw testManual sends distinguish address-validation failures from uncertain vehicle-command outcomes. A stop without a verified street location stays visible for correction or removal; a request rejected before dispatch does not pause future sends. See the October 1 error diagnosis.
The OCR release configuration uses eight CPUs, 8 GiB RAM, and two compute threads per engine. Native qualification, route-code filtering, and speed comparisons are recorded in the September 30 OCR report. The runtime and benchmark details below describe the September 28 release.
Image builds and publication are automated on pushes to main using the private
TrueNAS runner. GHCR receives latest and sha-<full-commit-sha> tags. See
GitHub Actions publishing for the runner and build setup.
Verified September 28, 2026. TeslaRouter runs 20260928-jlink, TrueNAS job
212741, from source commit 0fbb774d0cb2913e7e8151e09f4ed01c04dc861e and image
sha256:a62e3e8ff5390d2665030bd3306b7608c42bb0cde49acaad505e6bfd70df34bd.
The multi-stage image uses a Java 25 jlink runtime and runs the application as
UID/GID 10001, with 4 GiB memory, two CPUs, and a read-only root filesystem. The
existing route-data volume is retained and no host port is published. The JRE is
52% smaller; OCR dependencies and offline models still dominate image size.
iPhone HEIC/HEIF uploads convert automatically on the server. Select images from Safari's Photo Library or Files picker; no manual conversion or Shortcut is needed. Local conversion handles orientation before three local OCR engines run. The 12 MiB / 20 megapixel limits, TAP authorization, and CSRF protection remain in place. OCR subprocess environments exclude application credentials.
Native AMD TrueNAS qualification passed all nine synthetic cases (20 exact ordered rows) and both original HEICs (eight exact ordered rows). Every original row had majority support without unresolved disagreements. Peak memory was 1.12 GiB, with no OOM events or leftover jobs. Emulated amd64 timings from the Apple Silicon development machine are not used as production evidence.
The deployed browser flow at iPhone width passed a two-file uncached raw-HEIC upload (HTTP 200, 63.5 seconds), malformed-HEIC rejection (HTTP 400), busy-control restoration, duplicate-submit protection, and saved-session preservation. TAP consent loaded three authorized vehicles. No personal addresses were geocoded or vehicle commands sent. Physical iPhone hardware remains untested.
See release evidence and iPhone setup.