Custom Webhook Response: Send Your Own Status, Headers and Data
Reply to a webhook with data from the run: String, JSON or XML bodies, any HTTP status from 200 to 599, extra…
WordPress AI agent: build agents that reason and use tools
A WordPress AI agent in FlowMattic is a workflow step with instructions, a chat model and a toolbelt. Instead of following one fixed path, it reads the input, decides which tool to call, looks at the result and keeps going until it can answer. Everything runs inside the WordPress site you already have, on your own AI provider key.
Use an agent when the next step depends on reading something rather than on a fixed rule: a support message to classify, an order to judge, a lead to research. This guide walks through every setting on the AI Agent step, how the agent loop works, and seven workflows where an agent earns its place.
What is the AI Agent step?
The AI Agent is a built-in app called AI Agent by FlowMattic with one action, Execute Agent. You add it from the AI category of the app picker in the visual workflow builder. On the canvas it appears as a “Tools Agent” card with three connectors underneath: Chat Model (required), Memory and Tools. Each connector holds a small round node you click to configure.
An agent is different from a single AI completion. A completion answers once. An agent can call your workflow’s own actions as tools, many times in one run, and use what each call returns to decide its next move.
How a WordPress AI agent works, step by step
1. Write the instructions and the input
Click the agent card to open its panel. System Prompt defines the agent’s role and rules, for example “You classify support requests for Acme and tag them.” User Input / Message is required and is usually a dynamic tag from the trigger, such as {webhook1.message}. Typing @ in a field opens the dynamic tag picker.
2. Choose the chat model
Click Choose AI Model under the Chat Model connector. Pick a provider under Choose AI Provider, select the AI Connection that holds your key, then the AI Model. For most providers the model list is loaded live from the provider. Keys come from Connects, so you save one once and reuse it on every agent.
3. Attach tools
Click the + beside the Tools connector to add a tool node. Each tool is an app action: pick the app and its Action Event, then give it a Tool Name and a Tool Description so the model knows when to use it. On the Configure tab you fill the action’s fields the same way as a normal step, and Test Tool lets you check it before saving.
4. Let it loop
When the workflow runs, the agent sends the prompt and the tool list to the model. If the model asks for a tool, FlowMattic runs that action on your site and sends the result back. This repeats until the model answers without calling a tool, or until Max Iterations is reached. The final answer becomes the step’s output for the next step to use.
Providers and keys
The provider picker offers OpenAI, Anthropic, Gemini, Groq, Straico, OpenRouter, Perplexity, DeepSeek, Mistral AI, xAI (Grok) and AWS Bedrock, plus OpenAI-Compatible (custom URL). The last one takes a Base URL such as http://localhost:11434/v1 for Ollama or http://localhost:1234/v1 for LM Studio, an optional key connection, and a Model ID you can type or load from the server.
Requests go from your server straight to the provider with your key. FlowMattic adds no markup on tokens and there is no separate AI fee; you pay your provider at its own rates, or nothing with a local model. AWS Bedrock uses its own connection with a Bedrock API key or IAM access keys and a region.
Key options and settings
| Setting | Where | What it does |
|---|---|---|
| System Prompt | Agent panel | The agent’s role, rules and tone. |
| User Input / Message | Agent panel | Required. The message the agent works on, usually a dynamic tag. |
| Output Format | Agent panel | Plain Text or JSON. With JSON, a JSON Schema box sets the exact keys to return. |
| Max Tokens | Chat Model | Longest reply per call. Default 4096. |
| Temperature, Top P, Top K | Chat Model | Sampling controls. Defaults 0.7, 1 and 0. Values you change reach the providers that accept them; Top K goes to Claude and Gemini. |
| Frequency and Presence Penalty | Chat Model | Discourage repetition or new topics. Default 0. |
| Additional Parameters | Chat Model | JSON merged into the provider request for anything the form does not cover, such as reasoning or stop. null removes a parameter. |
| Max Iterations | Chat Model, Agent Limits | How many model turns the agent may take, 1 to 100. Default 10. |
| Tool Output Limit | Chat Model, Agent Limits | Cuts each tool result to this many characters before the model sees it. Empty means no limit. |
| Memory Type | Memory | No Memory (Stateless), Window Buffer Memory, WP Tables or WordPress Options. |
| Session Identifier | Memory | Groups memory by conversation. Leave empty for a new ID each run. |
| Include Execution Log / Request Body | Agent panel, Debug Options | Store tool calls, iterations and the raw request in Task History. Off by default. |
Memory across runs
Window Buffer Memory keeps the most recent messages (Window Size, default 10) for a day. WP Tables stores history in a table with session_id, role, content and timestamp columns. WordPress Options stores it under an option name that starts with fm_agent_memory_. To continue a conversation across runs, pass the same Session Identifier, for example a chat ID from the trigger. Memory keeps the user input and the final answer, not every tool call.
Who fills a tool’s fields
A tool field you fill, with a value or a dynamic tag, is fixed and the model cannot change it. A field you leave blank is described to the model, which fills it, so required fields are complete on the first call. If you want the model to fill only certain fields, start their values with AI: and an example, such as AI: [email protected].
What can be a tool
Any app action can be a tool, from WordPress and WP Tables to the API module and any of nearly 900 integrations. Flow-control and trigger modules are left out (Router, Branch, Filter, Iterator, Delay, Human in the Loop, Schedule, Webhook, Sub-workflow and the agent itself). Two tools deserve a mention: the MCP Client, which adds an external MCP server’s tools, and FlowMattic Browser’s Browse URL action, which lets the agent read live web pages.
Use cases
Support triage that tags and routes
A new ticket goes to an agent with tools to look up the customer in your CRM and search a WP Tables FAQ. It returns JSON with a category and urgency, and a Router sends urgent ones to Slack.
Help desk new ticket → AI Agent (CRM lookup, WP Tables) → Router → Slack · Email
Order review for a WooCommerce store
The agent reads the items, the total and the customer’s order history, then returns a priority the Router branches on. High-value or unusual orders reach the team; the rest get the standard confirmation.
WooCommerce new order → AI Agent (Output Format: JSON) → Router
Lead enrichment for a SaaS sales team
A sign-up arrives with only a name and an email. The agent calls the lookup tools you attached, fills company and role fields in a table and leaves a note on the contact for Amelia Okafor’s team.
Webhook (new sign-up) → AI Agent (API lookup, WP Tables update) → CRM add note
Research digest for an agency’s clients
Every Monday, an agent opens each client’s competitor pages with the Browse URL tool, notes what changed and writes one summary per client.
Schedule (weekly) → AI Agent (Browse URL) → Send email
Chat assistant with memory on any channel
Incoming messages from a chat channel arrive by webhook with a chat ID. The agent uses WP Tables memory keyed by that ID as the Session Identifier, so it remembers earlier turns, and the answer goes back to the channel.
Webhook (message) → AI Agent (WP Tables memory) → Telegram send message
Content drafting on publish
When a post is published, an agent writes social copy in your brand voice from the title and excerpt, then posts it to the team chat for review.
WordPress new post → AI Agent → Slack message
Private data on a local model
For internal records that should not leave your network, point the OpenAI-Compatible provider at an Ollama server and let the agent summarise rows from a table into a weekly report.
Schedule → AI Agent (OpenAI-Compatible, WP Tables get rows) → Send email
Tips and gotchas
- Describe tools well. The Tool Description is what the model reads when it decides which tool to call. Say what it does and when to use it.
- Set a Tool Output Limit for tools that can return a lot, such as a page of HTML or a large API response, so one result does not fill the context window.
- Max Iterations is a hard stop. If the agent reaches it, the step fails with “Maximum iterations reached. The agent may be stuck in a loop.” Raise the limit for genuinely multi-step jobs, or tighten the prompt.
- No automatic retries. A failed provider call fails the step. A failing tool reports its error back to the model, which can try another approach. A failed run can be re-executed from Task History.
- Tool calls run for real. If a tool sends an email or deletes a row, testing the agent does that too.
- Use JSON for branching. Output Format JSON with a JSON Schema gives later steps stable keys to map and a Router clean values to compare.
- One MCP server per agent. Add a single MCP Client tool per agent step, using None, Bearer Token or API Key authentication.
- Debug only when needed. Include Execution Log and Include Request Body increase what is stored in Task History, so switch them off once the agent works.
- A local model on another machine. Apart from localhost, WordPress only calls public hosts on ports 80, 443 and 8080, so an Ollama server at a 192.168.x address is refused until the site allows it with the
http_request_host_is_externalandhttp_allowed_safe_portsfilters. The error names both. - Text in, text out. The agent works on text. It does not take images or file attachments as input.
Related features
- MCP Client: give an agent the tools of any external MCP server.
- FlowMattic Browser: let an agent open and read web pages.
- Connects: save each AI provider key once and reuse it.
- Task History: see what every agent run received and returned.
FAQ
Do I pay FlowMattic for AI tokens?
No. The AI Agent runs on your own key, saved in Connects, and requests go from your server straight to the provider. You pay the provider at its own rates, or nothing with a local model.
Can a WordPress AI agent run on a local model?
Yes. Choose OpenAI-Compatible (custom URL), enter the Base URL of Ollama, LM Studio, vLLM or Azure OpenAI, and load or type the model ID. The key connection is optional, so a local server that needs no key works as is.
What stops an agent from looping forever?
Max Iterations on the Chat Model caps the model turns at 1 to 100, with 10 as the default. Tool Output Limit trims long tool results, which keeps each turn’s cost predictable.
How is an AI Agent different from Jaadu and AI Assistants?
An AI Agent is a step that runs each time its workflow runs. Jaadu is the copilot that builds workflows with you in the builder. AI Assistants are chat widgets visitors use on your site.
Can an agent remember earlier conversations?
Yes. Choose Window Buffer Memory, WP Tables or WordPress Options and pass the same Session Identifier on each run. Without memory, every run starts fresh.
Ready to give a workflow step a goal and a toolbelt? See FlowMattic plans and run AI agents on your own WordPress server with your own key.
