GPT Maker
AIBuild and operate AI customer service agents with GPT Maker, managing agents, training them, and conversing over the API. Agents keep support answers grounded in current documentation.
What This Integration Enables
GPT Maker is built around the assumption that a customer service agent will eventually need to stop and let a person take the conversation. Start Human Support is not a notification or a tag, it is a state change on the chat that stops the AI replying, and Stop Human Support is the explicit act that lets it start again. Nothing about it is inferred. A FlowRunner flow can therefore hold a conversation open under human control for as long as the case needs, on whichever channel the customer arrived through, and hand it back with a single call when the person is done.
The rest of the workspace is shaped the same way, as an operational system rather than a bot builder. Agents can be paused with Deactivate Agent and resumed with Activate Agent, which is a reversible act, next to a Delete Agent that is not. Credits are first class: Get Workspace Credits reports the balance and Get Agent Credits Spent attributes consumption to a single agent over a period, so a flow can check what a bulk training run will cost before it starts it. Interactions are separated from chats, because one customer chat can produce many support interactions as the same person comes back, and List Interaction Messages is the natural input to a summary or a quality score. Then there are intentions, which are the most powerful and most consequential part of the surface, and they are covered below because they deserve their own gate.
Without FlowRunner
With FlowRunner
Use Case Scenarios
A conversation handed to a person without the customer changing surface
A customer messaging over WhatsApp asks for a refund. A scheduled flow sweeping List Chats spots it, reads the exchange with List Chat Messages, and calls Start Human Support. The AI stops replying to that chat immediately. The agent opens a case in Zendesk with the transcript attached and posts the amount and the account into Slack. The rep answers the customer directly using Send Chat Message, which delivers on the customer's own channel on behalf of the account rather than asking the AI to compose anything. When the refund is settled, Stop Human Support returns the thread to the bot. The customer never left the conversation they started, and never had to explain it twice.
Knowledge kept current, with the cost checked first
Documentation is republished weekly. The flow calls Get Workspace Credits before it does anything, because GPT Maker bills training in credits and a large crawl on a low balance fails halfway through and leaves the agent partly retrained. With headroom confirmed, it runs List Trainings to see what the agent already knows, calls Create Training on the changed pages with the crawl options set, and supplies a callback URL so ingestion reports back when it finishes or fails rather than being assumed complete. Stale material comes out with Delete Training, which matters because only text trainings can be edited in place; everything else has to be removed and re-added.
An agent paused, not deleted
Support coverage ends at 8pm and the team would rather the bot say nothing than say something unsupervised overnight. The flow calls Deactivate Agent at close, which stops the agent answering on its connected channels while keeping its configuration, training and history intact, and Activate Agent in the morning. Deletion is a different question entirely. Delete Agent is permanent and the channels bound to that agent stop being served, so a flow proposing a retirement runs List Channels to name exactly which surfaces go dark, pulls Get Agent Credits Spent to show whether the agent is actually idle or just quiet this week, and puts both in front of the owner. Pausing is automated. Deleting is decided.
Human-in-Loop Highlight
The gate is on Create Intention, and specifically on intentions with automatic parameter and body generation turned on.
An intention is a standing rule that lets the agent call an external API when a customer's request matches its description. With automatic parameter and body generation enabled, the agent fills that request from the conversation instead of from a fixed template. Read that carefully, because it is not a normal write. Creating one intention is not one action taken once. It is an authorization: from that moment on, every matching conversation lets a language model compose an HTTP request against your systems using whatever the customer typed as its raw material. A single approved call can produce an unbounded number of future calls that nobody reviews individually. Delete Intention is likewise permanent, so removing one is not a rollback either, it is a second irreversible act.
So a flow that manages intentions never creates one on its own. It assembles the proposal and stops: "Proposing a new intention on the billing agent. Description: customer asks to change their plan. Method and URL: POST to the subscription service change-plan endpoint. Automatic parameter generation is ON, meaning the agent will fill the plan id and effective date from the conversation text. Currently 6 intentions are configured on this agent. Approve, approve with automatic generation off, or reject?" The owner picks. If they approve with automatic generation off, the intention ships with a fixed template and the model chooses when to call it but not what to send, which is a materially smaller grant. The agent has done the drafting, checked the existing intention list for overlap and named the endpoint plainly. What it will not do is grant a model standing permission to write to a production system on its own authority. That distinction, between doing the work and holding the decision that carries risk, is the entire point of human-in-the-loop as FlowRunner implements it.
Agent Capabilities
32 actionsWorkspace
2- List Workspaces Lists the workspaces on the account. A workspace owns agents, channels, chats and contacts, so its ID is the starting point for nearly every other operation here.
- Get Workspace Credits Returns the credit balance available to a workspace. GPT Maker bills conversations and training in credits, so this is the check that runs before a bulk campaign or a large training job.
Agents
8- List Agents Lists the AI agents in a workspace with paging and an optional name search. Each agent carries its own behaviour prompt, communication style, purpose and business details.
- Create Agent Creates a new agent in a workspace. The behaviour text is the system prompt, while communication style and purpose shape tone and objective, and the company details give the agent baseline knowledge before any training is added.
- Get Agent Retrieves one agent's full profile including behaviour prompt, communication style, purpose, company details and active state.
- Update Agent Updates an agent's profile, changing only the fields supplied. Used to revise the behaviour prompt, switch communication style or refresh the company details the agent relies on.
- Delete Agent Permanently removes an agent and its configuration, and the channels bound to it stop being served. This cannot be undone, which is why Deactivate Agent exists for reversible pauses.
- Activate Agent Turns an inactive agent back on so it resumes answering on its connected channels.
- Deactivate Agent Pauses an agent so it stops answering on its connected channels while keeping its configuration, training and history intact. The reversible alternative to deletion.
- Get Agent Credits Spent Returns how many credits an agent consumed over a period, narrowed by year, month and optionally day. Used to attribute cost per agent or to alert when one agent is burning the workspace balance.
Agent Settings and Webhooks
4- Get Agent Settings Returns an agent's operational settings: preferred LLM, time zone, whether human transfer and reminders are enabled, message splitting and grouping, emoji use, subject limiting and the per conversation message cap.
- Update Agent Settings Updates those settings, changing only the fields supplied. This is where the model behind the agent is chosen, human handover is enabled, replies are split or grouped, and a per conversation message cap and its penalty are set.
- Get Agent Webhooks Returns the webhook URLs configured on an agent for events such as a new message, a knowledge gap, a transfer to a human, and the first, start and finish of an interaction. These are plain URL fields on the agent rather than a subscription list.
- Update Agent Webhooks Sets those URLs. Each field is a single URL stored on the agent, so writing one replaces whatever was there before; read the current values first if only one is meant to change.
Conversing
2- Chat With Agent Sends a message to an agent over the API and returns its reply. The context ID identifies the end customer so the agent keeps one continuous conversation across calls. Text, an image URL or an audio URL can be sent, and a callback URL makes the call asynchronous.
- Add Context Message Writes a message into a conversation's history without the agent replying to it. Used after sending an outbound message from another tool so the agent knows what the customer was already told.
Training
3- List Trainings Lists an agent's training material of a given type with paging and an optional text search. Used to audit what an agent has actually been taught.
- Create Training Teaches an agent from raw text, a website, a video URL or a document URL, filling only the fields belonging to the chosen type. Website training can re-crawl on a schedule, and a callback URL reports when background ingestion finishes or fails.
- Delete Training Removes a piece of training material so the knowledge it carried is no longer used in replies. Only text trainings can be edited in place, so other types are removed and re-added.
Intentions
3- List Intentions Lists an agent's intentions, the rules that let it call an external API or follow special instructions when a customer asks for something specific. Supports paging and a text search over intention descriptions.
- Create Intention Adds an intention so the agent can call an external API, or follow a fixed set of instructions, when a request matches its description. For a webhook intention, turning on automatic parameter and body generation lets the agent fill the request from the conversation instead of a fixed template, which is a standing grant rather than a single action.
- Delete Intention Removes an intention so the agent stops calling that external API or following those instructions. This cannot be undone.
Chats and Human Handover
7- List Chats Lists the customer chats in a workspace, optionally narrowed to one agent, with paging and a text search. Each chat groups every message exchanged with one customer across their conversations.
- List Chat Messages Returns the messages in a chat, newest pages first, so a conversation can be summarised, audited or handed to another system.
- Send Chat Message Posts a message into an existing chat on the customer's own channel, sending text, an image, audio, video or a document. This delivers on behalf of the account rather than asking the agent to answer.
- Start Human Support Marks a chat as handled by a person, which stops the AI replying to it. Used when a flow detects frustration, a refund request or any case that needs human judgment.
- Stop Human Support Ends human handling of a chat so the agent answers again from the customer's next message onward. The explicit act that returns a thread to automation.
- Delete Chat Permanently removes a chat and its history from the workspace. Clear Chat Messages is the softer option when the chat itself should survive.
- Clear Chat Messages Deletes every message in a chat while keeping the chat, which also clears the context the agent was reasoning over. Used to reset a test conversation or honour a data deletion request.
Interactions and Channels
3- List Interactions Lists the support interactions in a workspace, optionally filtered to one agent. One chat can produce many interactions because a customer returns repeatedly, and each entry records when it started, when it finished and its status.
- List Interaction Messages Returns every message exchanged during one support interaction. The usual source for generating a summary, scoring sentiment or running quality analysis on a completed case.
- List Channels Lists the channels in a workspace, optionally filtered to one agent. Channels are the surfaces customers reach an agent through, such as WhatsApp, Telegram or the website widget.
Frequently Asked Questions
What can FlowRunner do with GPT Maker?
FlowRunner agents can run List Workspaces, Get Workspace Credits, and List Agents in GPT Maker, plus 29 more actions.
Does connecting GPT Maker to FlowRunner require OAuth?
No. GPT Maker connects to FlowRunner with an API key, no OAuth flow required.
Can GPT Maker trigger a FlowRunner workflow automatically?
GPT Maker doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with GPT Maker
$100 in credits. No card required. Connect in minutes.