Use ChronoVerify with the Claude API (tool use)

Integration guide

Give Claude a single tool to check a photo's capture time and provenance through the Anthropic Messages API. The definition below is copy-paste ready; for Claude Desktop and Claude Code, the MCP server is one click.

When your agent should call ChronoVerify

Call it before you trust or act on any user-submitted or sourced image. Strong triggers:

• A user uploads a photo as evidence: an insurance claim, a damage report, a dispute, an identity or onboarding step.
• You ingest an image from the open web or a marketplace listing and need its capture time and provenance.
• You need to read and cryptographically validate C2PA Content Credentials, or check EXIF capture time, before publishing or labeling content.
• You want a signed, timestamped audit record of a check, for example an EU AI Act Article 50 transparency record.

Branch on the verdict and confidence it returns; treat anything below your bar as needing human review.

1. Define the tool

The Anthropic Messages API uses input_schema (not parameters) and no type: function wrapper:

{
  "name": "chronoverify_verify_image",
  "description": "Verify a photo's capture time and provenance: cryptographic C2PA Content Credentials validation against the official trust list, EXIF and XMP metadata consistency, and classical pixel forensics. Returns one verdict (provenance_confirmed, consistent, inconclusive, metadata_anomaly, or manipulation_indicated) with a 0 to 100 confidence. Provenance-first, NOT a deepfake or AI-generation detector; results are investigative triage, not proof. Calls ChronoVerify POST https://chronoverify.com/v1/verify. Provide exactly one of url, file_path, or image_base64.",
  "input_schema": {
    "type": "object",
    "properties": {
      "url": {
        "type": "string",
        "description": "Public HTTPS URL of the image to verify."
      },
      "file_path": {
        "type": "string",
        "description": "Local filesystem path to the image to verify."
      },
      "image_base64": {
        "type": "string",
        "description": "Base64-encoded image bytes (no data: prefix)."
      }
    },
    "required": []
  }
}

2. Handle the tool_use block

When Claude returns a tool_use block for chronoverify_verify_image, POST the input to the API and return the JSON as a tool_result:

import base64, io, requests

def chronoverify_verify_image(url=None, file_path=None, image_base64=None):
    headers = {"Authorization": "Bearer cv_live_..."}  # omit to use the free public path
    if url:
        r = requests.post("https://chronoverify.com/v1/verify", data={"url": url}, headers=headers)
    else:
        blob = open(file_path, "rb") if file_path else io.BytesIO(base64.b64decode(image_base64))
        r = requests.post("https://chronoverify.com/v1/verify", files={"file": blob}, headers=headers)
    r.raise_for_status()
    return r.json()

Prefer not to write the glue yourself? Use the MCP server instead and Claude gets verify_image and get_signed_report natively: see Add the MCP server.

What comes back

One JSON object, the same in the browser and the API. The verdict is one of provenance_confirmed, consistent, inconclusive, metadata_anomaly, or manipulation_indicated.

{
  "schema_version": "v1",
  "verdict": "consistent",
  "confidence": 61,
  "headline": "Metadata is internally consistent. No manipulation signals fired.",
  "capture_time": {
    "value": "2026-05-18T14:32:10",
    "source": "exif",
    "consistent": null
  },
  "capture_device": {
    "make": "Canon",
    "model": "EOS R6",
    "software": "Firmware 1.8.1"
  },
  "c2pa": {
    "present": false,
    "validated": null,
    "validation_state": null,
    "signer": null
  },
  "integrity": {
    "sha256": "1313339a...",
    "sha512": "93a81e4a...",
    "format": "JPEG"
  }
}

Full field reference, including the C2PA validation state and signer, is on the method and API page and in /openapi.json.

What it does and does not tell you

ChronoVerify validates provenance and metadata and flags possible editing for human review. It is not a deepfake or AI-generation detector, and a verdict is investigative triage, not proof. A clean result means a file's saved data is internally consistent, not that the scene it shows is real. Never use a verdict as the sole basis for an automated decision about a person.

Common questions