How it works
What every agent comes with, how it connects to your tools, the three ways to use it and where your data stays.
Configuration lives inside the template. Each kind of agent has its own section group.
Goal, do/don't guidelines with a flow graph, output schema builder, rules as sentences, tools and models across Claude, Gemini and OpenAI, skills library, knowledge, data sources, allow-listed web search, ready-made report layout, test files.
87 difficult samples, with the report and the CSV shown before you copy anything.
Live timeline, fields in Italian, confidence, sources, rule checks, approvals, JSON/CSV/report downloads, evals.
27 prebuilt workflows, knowledge ingestion, 11 kinds of read-only data sources, phone and WhatsApp channels.
Local runner on the client's machine, no-retention mode, test-set grading, EU/US data-region policy.
Copy, edit, version, remove. Or describe the change and the copilot edits the agent. 3 console languages, pay-per-use credits with spend caps.
The agent reads the documents where they already are, consults the tables you let it read, and writes the results back beside them — on our servers or on the client's own machine.
Point the agent at the folder the documents already arrive in — SFTP or FTP, S3, Azure, a shared folder. It reads each one from there and writes the results back next to them: the CSV, the JSON and the report PDF into a processed folder, one run at a time. The inbox is read, never emptied.
fatture/inbox → fatture/processed
A read-only role on the ERP's Postgres or MySQL is all it needs. The tables the agent may consult are saved once — the open purchase orders, say — and read during the run to check what it extracted. The agent never writes to your database.
ordini_acquisto · stato = 'aperto'
The same agent runs on the client's own machine with afb-runner, on a local folder and the local database. The documents never leave the room; only the counts and the seat come back to the platform. A signed definition keeps it valid for a fixed period.
afb-runner doctor → afb-runner ui
Copy a template, configure it, run it, read the trace. 3 languages.
Start freeSDK and webhooks. The same agents from your own systems, the same trace.
Read the API referenceBranded, on your own domain, installable as a PWA for your team or your customers.
Read the guideOur templates run on Claude and Gemini by default, and on a paid plan an agent can be set to another model. Local runner, no-retention mode, an EU/US data-region setting that limits which models an agent can be saved with, test-set grading. The trace names who processed each step, where, and what it cost.
Claude Opus 5 · Claude Sonnet 5 · Claude Haiku 4.5
template default
Gemini 2.5 Flash
template default
GPT-5 mini
via OpenRouter
Mistral Large
not connected yet · hosted in France · 30-day retention
Qwen 3 · open weights
not connected yet · hosted on Scaleway · zero retention