Case Study · AI Agents
An MCP server that turns a sentence into a working chat flow
Our client's SaaS platform lets businesses build chat flows node by node. Building big flows was slow and needed expert users. AMT added a Model Context Protocol server that generates a full flow from a plain-language prompt and suggests the next node while users build by hand.
Schedule Technical DiscoveryYour first conversation will be with a senior technical partner, not a salesperson.
Published
At a glance
- Client
- Company running an in-house SaaS automation platform with a node-based chat flow builder
- What AMT built
- An MCP (Model Context Protocol) server connected to the platform's backend, workflow metadata and AI models
- AI models
- OpenAI models and Anthropic models, selected per task
- Stack
- Angular, Node.js, Python, MongoDB, OpenAI API, Anthropic API
- Result
- Faster flow creation, less manual node configuration, usable by technical and non-technical users
The short version
Users can now type "Create a customer support flow for refund requests" and get a structured chat flow with the right nodes and connections. While they build manually, the server reads the flow so far and recommends the next logical node. The MCP server sits between the SaaS backend, the platform's node configurations and two AI providers, and picks whichever model suits the task.
The problem
The platform already supported chat flows for customer support, onboarding, notifications and task automation. Building them was the hard part:
- Users found it difficult to pick the right sequence of nodes.
- The builder gave little guidance while a flow was being built.
- Configuring each node by hand took time.
- Large conversational workflows got complex fast, so only experienced users built them well.
What AMT built
An MCP server as the AI layer
AMT built a Model Context Protocol server that works as an intelligent service layer. It talks to three things:
- The SaaS platform backend
- Workflow metadata and node configurations
- AI models from OpenAI and Anthropic
Because the server has the platform's real context, its suggestions match how flows actually work on that platform, not generic chatbot logic.
Whole flows from a prompt
Users describe the workflow they want in plain language, for example:
- "Create a customer support flow for refund requests"
- "Build a chatbot flow for booking confirmations"
The MCP server turns the prompt into a structured chat flow with nodes and connections the user can then edit.
Next-node suggestions while building
When users build by hand, the server analyses the current flow and recommends the next node. Suggestions draw on:
- Nodes already in the flow
- Node dependencies
- Historical usage patterns
- The flow's logic
Two model providers, picked per task
- OpenAI models handle reasoning, workflow generation and intent understanding.
- Anthropic models handle structured responses and contextual recommendations.
The server chooses the most suitable model for each task, which also gives the platform a fallback if one provider has problems.
Results
- Faster chat flow creation
- Intelligent node recommendations inside the builder
- Less manual configuration
- More accurate workflow designs
- A builder that non-technical users can use, not just experts
Technology stack
| Layer | Technology |
|---|---|
| Front end | Angular, HTML, CSS |
| Back end | Node.js, Python |
| AI integration | OpenAI API, Anthropic API |
| AI layer | MCP server (Model Context Protocol implementation) |
| Database | MongoDB |
Before you build something similar
- What context does the model need? An MCP server is only as useful as the product data it can see. List the metadata, configurations and history it should read.
- What should the AI be allowed to change? Generating a draft flow the user reviews is very different from editing live flows. Decide that boundary first.
- One model or several? Routing tasks to different providers can improve quality and resilience, but you have to test each route.
- How will you measure success? Time to build a flow and how much users edit the AI's draft are both good signals.