About ChatFrom

ChatFrom is an AI chatbot platform that answers from your own documents and website, and acts through your own tools, for businesses that answer their customers in chat, with or without a support department.

Last updated 4 November 2026

What ChatFrom does

Answers from your knowledge

You add PDFs, Markdown, text, questions and answers, pages from Notion, documents from Google Docs, or the address of your website, and ChatFrom indexes them for that chatbot alone. Every answer names the pages it came from, so a visitor can check it and you can see where a wrong answer began.

Actions in your own product

A chatbot can call your API through a tool server you run (MCP), so it looks up an order, creates a project or books a call instead of only describing how to. On your site it carries the signed-in user's token, so it acts as that person and no further.

The services you already run on

Connected to the services a business runs on, a chatbot acts for the customer it is talking to:

• Shopify: where an order is, with the tracking link, once the customer gives the order number and the email it was placed with, and a return request, which you approve or decline in Shopify.
• Stripe: a customer's subscription, cancelled at the end of the paid period or resumed, and a link to update their card, once they prove their email with a code.
• Google Sheets: every lead, booking or request as a new row under your sheet's own columns.
• HubSpot: every lead saved as a contact in your CRM, with a note of what they asked.
• Cal.com and Calendly: your free times offered in the chat, and the one the visitor picks booked.

Anything else with an API is reached through a custom HTTP action or a tool server (MCP), Zapier's among them, which reaches thousands of apps.

A visual editor for customer flows

You draw how each kind of conversation should go, step by step on a canvas: search the knowledge, ask the visitor a question, branch on the answer, call a tool, hand over to a person. That is how a refund request and a pricing question take different paths, one for every scenario, without anyone writing code.

Channels

The same chatbot answers on your website, through a share link, in WhatsApp, Telegram, Slack, Discord, a Crisp chatbox, and by email when you forward your support address to it. With the iOS SDK it also answers inside your own iPhone and iPad app. Every conversation lands in one log, whichever channel it came from.

Conversations, analytics and evaluations

Every conversation from every channel is in one log, and you can filter it to the ones the chatbot could not answer, the answers visitors rated badly and the ones handed to your team. Each answer shows its model, its tokens and its cost. Analytics shows the last 7, 30 or 90 days: conversations each day with the share left unanswered, what they cost, and how visitors rated the answers.

A weak answer becomes, with Fix this, a question-and-answer entry the chatbot uses next time. Before you change anything, an evaluation asks your chatbot a set of test questions and a judge model from another AI company scores each answer, or a RAGAS run scores how well it found your knowledge. Answers can also be traced in your own Langfuse project.

What makes ChatFrom different

The engine is open source

The part that writes the answers is a separate service, published under Apache 2.0 at github.com/rouzbeh-abadi/chatbot-engine. You can read exactly how retrieval, tool calling and workflows work, run it yourself, or keep using ours. Most chatbot platforms keep that part closed.

Your choice of model, counted in credits

Models come through OpenRouter, so OpenAI, Anthropic, Google and Mistral are one setting apart. Each answer is counted in credits from your plan, and the Model page shows how many credits an answer uses on each model, so you can compare before you switch, rather than finding out from a monthly bill.

Your tools, not our integrations directory

Instead of waiting for a platform to support your system, you expose the actions you already have as a tool server and allow the chatbot the ones you choose, per chatbot. The chatbot never sees anything you did not list.

Paid by what it uses

Plans are monthly, with no minimum term and no per-conversation or per-resolution fee: a plan is an allowance of answer credits and knowledge, and the free Hobby plan needs no card. You can also bring your own OpenRouter key for evaluations, so testing costs you provider prices and nothing more.

Built and hosted in Germany

The company, the servers and the database are in Germany: the site, the app and the engine run in Frankfurt, and the database is there too. The data processing agreement is part of the terms you accept when you sign up, so there is nothing extra to sign, and every sub-processor is named in it.

How good are the answers?

We test the engine on questions whose right answers we know. The tests call real models, so they are run by hand, and the table shows the last run of each. The golden set and RAGAS were run in September 2026, on engine 0.1.18 for the golden set and 0.1.0 for RAGAS, with answers from GPT-5 mini, the default model until 16 October 2026.

The Shopify customers were run again on 22 October 2026, on engine 0.1.39 and GPT-6 Luna, the default now. The RAGAS evaluation is part of the open-source engine.

The golden set holds one question the shop's pages do not answer, one in German and one aimed at an instruction planted in a page. It caught two real problems, both fixed.

• Keyword search let passages that only shared a common word push out the one closest in meaning. Counting only the keyword search's three best matches fixed it.
• German questions got answers that kept the English wording of the pages. A rule and a worked example took fully German answers from 8 of 10 to 20 of 20.

RAGAS scores answer relevancy lowest because it counts an honest "the documents do not say" as zero, which is the answer we want. You can run both kinds of test on your own chatbot under Evaluate, and what an answer costs on each model is on the pricing page.

Who uses ChatFrom

• Small SaaS teams, roughly 1 to 20 people, who answer support email themselves and have no support department.
• Agencies and studios building a chatbot for a client, who want the answers grounded in the client's own documents.
• Online shops that answer the same questions about delivery, returns and orders every day.
• Product teams who want an assistant inside their signed-in app, acting for the user through their own API.
• Developers who would rather read the engine's code than trust a black box.

The team behind ChatFrom

ChatFrom is made by DeadPixel Studio UG (haftungsbeschränkt), a studio in Berlin run by Rouzbeh Abadi, its founder and managing director. Rouzbeh is a software engineer who has spent years building web products, and writes the code, the docs and this page himself; his work is at github.com/rouzbeh-abadi.

The studio built ChatFrom while making another product, Lighthouse, that needed a support chatbot able to look at a customer's own account. Nothing on the market both answered from our documents and acted through our API without a rebuild, so we built the engine first, put a product around it, and now run our own support on it.

It is a small team: the founder, with help from contractors on design and content. You will not be handed to an account manager, because there is not one.

How ChatFrom works with you

• Getting started takes a few minutes and no call: sign up, add a document or your website address, set how the chatbot should talk, and put one script tag on your site.
• Who you deal with: the founder. Questions about the product, a bug or an invoice all reach the same person.
• How to reach us: the contact form or email at rouzbeh@deadpixelstudio.io. We answer within two working days, usually sooner. ChatFrom is on X as @chatfrom_io, and the founder's account, @JustRouzbeh, is open too.
• What we need from you to set a chatbot up: your documents or a website to crawl, and, if it should act for signed-in users, a tool server it may call.
• Changes and upgrades: a plan is monthly and ends when you say so, with no notice period. Deleting a chatbot deletes what it knew with it.

Key facts

Frequently asked questions