Graph Agent
Experimental
Intake and three-way match run exactly as certified in Expert Review, as configured live in the Visual DAG — no LLM rewrites a rule. Duplicate/anomaly interpretation and approval reasoning are LLM-backed (Claude): the model can explain and draft an override note, but it cannot override a hard block or flip a pass/fail rule. Runs are recorded into Invoice Activity like any other run, tagged as coming from this experimental view. Anything needing human review is handed off to Observability, where it can be approved, overridden, or rejected.
Following the Visual DAG
Every run loads the live Visual DAG configuration and gates its behavior on it — a branch switched off there actually changes what happens here, not just how it is drawn.
All branches enabled — running with the fully certified default configuration.
Open Visual DAG Testing utility
Clears invoice_review (the record of every invoice number this agent has already processed) so you can re-run the same invoice without tripping the duplicate check or other checks that look at prior runs.
Anthropic API key
Used directly from your browser to call Claude and never sent anywhere else. Held only for this browser tab (sessionStorage) — never written to disk or synced. Use a key you're comfortable pasting into a client-side prototype, not a production key.
Invoice to run
Same fields as New Invoice Review — only enabled channels are shown. Fill them in by hand, or upload an invoice document to auto-fill them.
Enter an API key above to upload a document.
Try PO-58910 (partial delivery), PO-59004 (no delivery receipt, new vendor), or invoice number INV-77341 for Meridian Office Supplies (duplicate) to see different branches.
Enter an API key to run.