Action Verification

When a workflow takes a UI action, farscry diff returns what changed locally, $0, no cloud round-trip.

The core pattern

farscry extract before.png -o before.vasp

farscry diff before.png after.png

MCP server setup

The recommended integration uses farscry as an MCP server so your workflow can call farscry_extract and farscry_diff directly.

Installation

npm install -g farscry

Or: pip install farscry

Configure MCP settings

Run farscry setup to auto-detect your agent and get the config snippet to paste.

Or add manually to your agent’s MCP config:

{
  "mcpServers": {
    "farscry": {
      "command": "farscry",
      "args": ["serve", "--mcp"]
    }
  }
}

The MCP host starts the server automatically. The server keeps OCR engines warm across calls.

What the workflow sees

Before farscry

[a workflow takes action on UI]
[workflow sends a new screenshot to remote processing: $0.0047, typically 2-5s (cloud round-trip)]
[workflow receives a full-screen description]
[workflow extracts what changed manually]

With farscry

[a workflow takes action on UI]
[farscry diff before.png after.png: $0, locally, no cloud round-trip]

appeared:
  - button "Retry" at (300,420)  enabled:true
changed:
  - button "Submit" → "Processing..."  disabled:true
removed:
  - spinner at (450,200)

Example: verifying a payment flow

farscry extract payment_form.png --affordances
affordances:
- type:  input "Card Number"   current: ""
- click: "Submit Payment"      enabled: true

cp payment_form.png before.png

farscry diff before.png after_submit.png
appeared:
  - spinner at (300,200)
  - text "Processing payment..."
changed:
  - button "Submit Payment" → disabled:true

farscry diff after_submit.png after_complete.png
appeared:
  - badge "Payment successful"  state:success
removed:
  - spinner

Performance comparison

Approach	Latency	Cost	Offline
farscry diff	locally, no cloud	$0	Yes
remote screenshot processing	~2-5s (unconfirmed exact)	$0.0047/img	No

Over a 20-step automation session with 10 screenshot verifications: farscry saves several seconds and ~$0.047 per run.

Loop detection

farscry’s state_id (perceptual hash) enables loop detection in automation workflows:

farscry extract screen.png --context
Payment settings, Save available

For state_id, run farscry extract screen.png without --context.

If the same state_id appears twice, the workflow is in a loop, the action had no effect.

farscry supports this to detect when an action had no effect and bail out early.