1,914,490 street photos of Tokyo, read by AI

The street details maps leave out, searchable in words.

Sidewalk widths, signs, planting, the age of buildings. AI described street photos of Tokyo taken by citizens, one by one, so you can look up how places actually are, by location or by words. Descriptions are in Japanese.

Try 懐かしい路地nostalgic alley 涼しげな場所cool, shady place 📍 Near me
See all of Tokyo on the map →Use the API or AI →
  • No key or sign-up
  • Works with AI agents (MCP)
  • Words under CC BY 4.0 (no credit needed in answers)
From the live APIone photo → words
© Mapillary contributors · CC BY-SA 4.0 · resized

    Translated from the Japanese record. The API returns the original Japanese text.

    GET /v1/scenes/1580655655467339 ↗
    1,914,490street photos described
    8,297streets and town blocks written up
    3,618changes confirmed in photosChecked by a separate AI session.
    2014–2026capture years (99.8% of photos)

    All of Tokyo (the 23 wards over several years; Tama and the islands mostly the latest year) · release 2026-09-13-r2

    Use it three ways

    On screen, by API,
    from AI agents.

    All three run on the same public data. No sign-up, no API key.

    01Look on screen

    Three screens that open in any browser.

    Mapphotos, streets, blocks, changes Search scenesJapanese interface 3D citywords written on Tokyo's 3D city model (Japanese)

    Screens reduced. Photos © Mapillary contributors (CC BY-SA 4.0) · Map: OpenFreeMap © OpenMapTiles © OpenStreetMap contributors (ODbL) · 3D city model: Project PLATEAU (MLIT), processed

    02Query the API

    Send a location and get descriptions of photos, streets, blocks and changes. REST and OpenAPI.

    curl "https://michiyomi.dev/v1/scenes/nearby?lat=35.6717&lon=139.7647&radius_m=150"
    // nearby photos (excerpt; descriptions are in Japanese) { "capture_year": 2019, "ward": "Chuo", "summary": "sidewalk: left=yes (separated, 5m), right=yes (separated, 4m), tactile paving …" }

    API reference · OpenAPI · up to 300 requests a minute

    03Use from AI

    Paste this line into Claude, Codex or another agent. It reads the setup guide and connects itself (MCP).

    Read https://michiyomi.dev/agent-setup.md and set up the michiyomi MCP server.
    claude mcp add --transport http michiyomi https://michiyomi.dev/mcp

    Use from AI · /llms.txt · no uptime guarantee (SLA)

    1. 1 On the map
    2. 2 In words
    3. 3 In 3D

    From one photo to streets,
    blocks and years.

    The same readings, gathered at four scales. These examples from Ginza and Shinjuku come straight from the live API. All of them sit on one map and lead back to the original photos. (Texts below are translated from the Japanese records.)

    Pointone photo
    A narrow street in Shinjuku with wooden shops and parked motorbikes
    MapillaryPhoto: a Mapillary contributor, CC BY-SA 4.0, cropped and resized

    A narrow street in Shinjuku 2025

    Wooden shops and houses on both sides. The left shoulder is about 1.5 m wide, with two motorbikes parked on it.

    On the map →/v1/scenes
    Areatown block profile
    The map: the outline of the Ginza 4-chome town block
    Map: OpenFreeMap © OpenMapTiles © OpenStreetMap contributors (ODbL) · Block outlines: 2020 Census (e-Stat), processed

    Ginza 4-chome (Chuo)

    Glass-fronted high-rise offices and shops line wide arterial roads, with brand signs and broad sidewalks.

    From 1,001 photos, 2010–2026

    Read the profile →/v1/areas
    Timechanges over years
    The road in 2019: the left edge of the roadway is plain grey, with white markings only
    2019
    The same road in 2026, with a blue band along the left edge
    2026
    MapillaryPhotos: Mapillary contributors (image pages: 2019, 2026), CC BY-SA 4.0, cropped and resized

    A main road in Yotsuya, Shinjuku

    In the 2026 photo, a continuous blue band marking where bicycles ride runs along the left edge of the roadway, which was plain grey in 2019.

    Only changes that survived a refutation attempt in a separate AI session are kept

    On the map →/v1/changes

    Search street scenes
    in words.

    Write something like 懐かしい路地 (a nostalgic alley). It searches the text AI wrote about each street photo and returns places that fit, with photos. Search in Japanese.

    The scene search screen, in Japanese: a search for 'nostalgic alley' returns photos of narrow alleys lined with old houses and shops, each with its year, place and why it fits
    Results for 懐かしい路地 (nostalgic alley), one search from October 2026. Each candidate shows its capture year, place and why it fits; view them on the map or save them as GeoJSON.MapillaryPhotos: Mapillary contributors (image pages: left, middle, right), CC BY-SA 4.0 · screen reduced
    1. 1 Gather up to 30 descriptions closest in meaning
    2. 2 Check AI reads each and scores the fit
    3. 3 Rank passing photos become candidates

    The AI judges from the text written about each photo, not from the photo itself. Descriptions show each street as it was when photographed.

    Search scenes (Japanese)Near me (Japanese)
    Try映画のワンシーンみたいlike a scene from a film秘密基地みたいなところlike a secret hideout

    A 3D city
    made of words.

    What AI read in street photos, written onto Tokyo's buildings and streets in PLATEAU, Japan's national 3D city model. The interface is in Japanese.

    The 3D city: buildings and streets around an intersection in Shinbashi, drawn with the Japanese words read from street photos (Japanese interface) What you see when you tap a building: its type (a high-rise office building), how many photos and from which years, how many descriptions, and the photos themselves (Japanese interface)
    Shinbashi up close. The characters on walls and streets are words read from photos; the overlaid panel is what you see when you tap a building.3D city model: Project PLATEAU (MLIT), processed · Photos © Mapillary contributors (CC BY-SA 4.0) · screen reduced
    1. 1 Pick one of 96 places and look around
    2. 2 Tap to see words and photos
    3. 3 Search steps, vending machines, tactile paving…
    726,767buildings, streets and more with words
    about 83.66Mcharacters on walls and streets
    Open the 3D city (Japanese)

    Watch the Shinbashi intro (about 40 s, Japanese) →

    Grand Prix, PLATEAU Hack Challenge 2026 in TOKYO (michiyomi × LOOVIC joint team; page in Japanese)

    Built to be checked

    AI readings,
    in a form you can check.

    AI misreads. So we keep who wrote what, when, and from which photo, and return how sure each reading is.

    A label on everything

    Every record carries its capture year, the generation of AI that read it, and the data release. Image IDs lead back to the original photos.

    Estimates say so

    Widths and lane counts are estimates with a confidence level. "None", "unknown" and "out of frame" are kept apart, and nothing unseen is filled in.

    Stability, measured

    Agreement rate when the same photo was read twice (36,479 photos).

    • Light and weatherover 95%
    • Road type85–90%
    • Sidewalk present75%
    • Exact width57%

    Only changes that held up

    Changes found in photos taken years apart were challenged in a separate AI session. Only the 3,618 that held up are published.

    A separate AI compared readings against photos: 73.3% judged correct (83.8% including minor errors, all fields of 2,468 uniformly sampled photos). Agreement is stability, not accuracy. Numbers are estimates from images, not measurements. Use it to narrow candidates or understand trends. Check the photo or the site before making an individual decision. More: How it works (JA) · Accuracy check (EN) · docs.

    Open

    The API, the data
    and the model are open.

    The raw material is photos that citizens took and shared. The readings, and the model that reads, are open to everyone in the same way.

    No sign-up · read-only

    API and MCP

    A REST API for points, streets, town blocks and changes, and an MCP server for AI agents.

    Open the docs →     CC BY 4.0

    Dataset

    A public snapshot in Parquet for research and analysis. It may lag behind the API.

    Hugging Face ↗     CC BY 4.0 · PLATEAU

    Linked to the 3D city model

    A table linking street observations to 727,000 PLATEAU buildings, roads and other features by feature ID, with a tool that writes them into your own CityGML.

    Hugging Face ↗     Apache-2.0 · experimentalOddEye

    Our street-reading model. It writes one street photo as what it sees and what it infers. Median 27.5 s per photo on one RTX 5090.

    observationsinterpretationsSee the model ↗

    The public data on this site was made by our production pipeline. OddEye is a separate model for the next, richer record (13 fields); we have not changed how the public data is made.

    News

    News