The context layer for the physical world

• The Space Document

  Areas with type and condition; items with location, quantity, brand and the names people use for them; issues with severity and a next step; key locations; how-to steps; facts; and what the video did not cover.

  JSON · versioned

  {
    "id": "itm_32",
    "name": "Portable fan heater",
    "category": "climate",
    "location": "On the floor below the left side of the vanity",
    "condition": {
      "code": "damaged",
      "notes": "The power cable is visibly loose."
    },
    "issue_ids": ["iss_2"],
    "evidence": [
      { "frame_id": "frame-28", "timestamp_ms": 28087 },
      { "frame_id": "frame-30", "timestamp_ms": 30058 }
    ]
  }

• Evidence frames

  Every record cites the frames of the original video it was seen in, through short-lived signed links. Each item, issue and key location is then checked again on up to four of them by a second model: on photo, unconfirmed, or from the narration when it cites none.

  Signed URLs

  • 0:00
  • 0:15
  • 0:24
  • 0:30

• Answers with citations

  A short answer from the space’s own records, with its sources and photos, or “not documented”. Never a guess from a typical home.

  POST /ask

• Search across spaces

  “Where did I see a ladder?” finds the record in any of your spaces, in any language: semantic search with a reranker.

  GET /search

• Compare two scans

  Check-in against check-out: items added, moved, recounted; gone only when a frame shows the empty place; issues new or fixed.

  GET /diff

• Agent Skill and Markdown

  Export a space as a skill folder any agent loads on demand, or as one Markdown page to paste into a ticket or a prompt.

  SKILL.md

• Live events and webhooks

  Follow a scan as a stream of events; receive signed scan.completed and scan.failed on your server, with retries.

  SSE · HMAC

• Guided capture

  What to film for each type of space, from flats and offices to cars, boats, campers and appliances; each frame checked in about half a second: usable or not, which key locations it shows.

  POST /frames/check

1. 01

   Film

   Walk through with a phone: a video of up to 10 minutes and 4 GB, or up to 120 photos. Say what the camera cannot see: the Wi-Fi, the bins, where the fuse box is.

2. 02

   Scan

   One call from the SDK, the CLI, an assistant or the console. The analysis runs on our side, with no ffmpeg on yours, and reports every stage as it happens.

3. 03

   Your agent knows it

   Read the Space Document, search it, ask it with citations, compare it with the next scan, or hand it to Claude as a skill.

• Test

  Free

  Build the whole integration before you pay anything.

  • The demo space: two real scans of a one-bedroom apartment
  • Free test keys, no card needed
  • Every endpoint, the SDK and the CLI
  • Scans complete instantly and nothing is charged
  Get a free test key

• Live

  First scan free

  $1.00per scan

  Your own spaces, from your own videos.

  • Your first live scan is free (one per card)
  • Prepaid balance from $10, by card through Stripe
  • Live scans need a card on file; saving it charges nothing
  • Failed scans are refunded to the balance
  • Reading, search and questions included
  • AI agents can pay by themselves in USDC on Base or Solana (x402)
  Open the console

• Higher volume

  Let’s talk

  Many spaces, many scans, or limits above the defaults.

  • Higher limits on scans, search and questions
  • A direct line to the team that builds the API
  Contact us

Pay as you go, no subscription, the balance never expires.

Full pricing and questions

• Redaction by scope

  Wi-Fi passwords, door and alarm codes stay hidden from every key without the spaces:sensitive scope: in documents, search results, answers and exports.

• Instructions in videos are data

  A sign or a sentence on camera that tries to give orders to an AI is withheld before any agent reads it and before the models that write the document, and flagged in the document.

• Files stored in the EU

  Videos and frames are stored on Cloudflare R2 in the EU jurisdiction, and the source video is deleted once its frames are extracted.

• Signed webhooks

  HMAC-SHA256 over the timestamp and the body, sent only to public HTTPS hosts, never redirected, retried with backoff.

• Evidence from the original video

  Every record points to the frames it was seen in, served through short-lived signed links.

• Keys you control

  Hashed, shown once, scoped, with an optional monthly spending cap. Each MCP connection gets its own key; the audit log keeps who did what for 400 days.

Security and privacy

• 14 October 2026 · 6 min read

  Campervan and boat handover inventory from one video

  What to film when a campervan or a boat changes hands, what the API records (VIN plate, gas, battery, seacocks) and how to compare the handover with the return.

• 14 October 2026 · 6 min read

  RoomPlan alternative without LiDAR: what is in the room

  Apple RoomPlan maps walls and furniture boxes on LiDAR devices. To know what is in a room, where and in what state, from any video, here is the other half.

• 14 October 2026 · 7 min read

  Vision model hallucinations: check each claim on its photo

  A vision model can name an object its photo does not show. How Space Context asks every record again on its own frames, and what 38 of 38 and 0 of 227 mean.