Skip to content
Hardik Zinzuvadiya
Let's connect

product

Frappe Copilot

An AI agent inside ERPNext that answers from live data and drafts records for human approval.

Watch the demoMore info ↗

Overview

Frappe Copilot is an AI agent embedded in the ERPNext Desk. It answers questions from live, permission-checked data, turns supplier invoices and receipts into draft documents, and puts every write in front of a person before it lands. Bring your own model; the data stays on your site.

28tools the agent can call
6LLM providers with automatic failover

Features

Agent

  • Chat inside the Desk that plans, calls tools, reads results and repeats until done
  • 28 tools for reading, analysing, drafting, extracting and navigating
  • Skills are records: write one in the Desk and the next message already uses it
  • Each skill can carry knowledge sources: your notes, a documentation URL, or a scan of a DocType
  • Book this invoice: reads the PDF, matches supplier and items, drafts, asks for approval
  • Mention any document with @ to pull it into the prompt

Routing, caching and failover

  • An intent router reads the message first and picks the smallest tool and skill set the executor will see
  • It ships off, and returning nothing is a supported answer, so an install behaves the same until you turn it on
  • Greetings and acknowledgements never reach it, so the cheapest messages cost no routing call
  • Prompt caching pins the stable layers: identity and moderation, then the skills block, then the tool catalogue
  • Long conversations compact themselves past 60% of the token budget; the last eight messages stay verbatim
  • The compacted block keeps figures bound to the tool and filter that produced them, so the next turn recovers state instead of re-reading prose
  • Rate limits and outages move to the next provider mid-run; an auth failure skips that provider outright

Answers

  • Records, aggregations, full-text search, document link tracing and saved reports
  • Sentences, sortable tables or charts, streamed over SSE as real UI components
  • No vector index to maintain and no stale copy of the database
  • Compact tool results use 40-60% fewer tokens than JSON

Extraction

  • Digital and scanned PDFs, photos and FatturaPA e-invoices become Purchase Invoices
  • Expense Claims from receipts through the same review card
  • Supplier and line-item matching, totals checks, confidence scores
  • Uncertain lines show ranked candidates; nothing is silently guessed

Control

  • Runs as the logged-in user: if you cannot open a record, Copilot cannot either
  • PrivacyGate masks sensitive fields before any prompt leaves the server
  • Every write is a draft card; nothing changes until you approve
  • A dry run validates each draft first, so errors show before you confirm
  • Analysis runs in a sandbox: restricted Python, masked output, and a rollback even on success
  • Admin centre: setup wizard, usage, traces, per-user quotas
  • Prompts are versioned records, moved from draft to canary to active to retired

Technical details

Built with
Python, Frappe v15/v16, React, SSE streaming
Type
product

Have something slow, manual, or fragile in production?

Tell me what it is. I reply within two working days, with questions.

Let's connect