MazeGraph builds a request for each reply and sends it through WordPress’s AI Client to the provider selected for that session. It includes the message and the context needed to answer. The exact contents vary with the conversation, the selected assistant, enabled tools, and your context settings. Before a person’s first message, MazeGraph shows a data notice and asks for consent.
Information that may be sent
| Information | When it may be included |
|---|---|
| Your message | The text you send in the chat or a Content Studio stage. |
| Conversation history | Relevant earlier messages that fit the context strategy configured for long conversations. The full history is not necessarily sent each time. |
| Instructions and site context | MazeGraph’s assistant instructions, site or personal instructions, and basic context such as the site name and URL, the signed-in person’s display name, or the current date. |
| Tool descriptions | Descriptions of tools available to the selected assistant, so the provider can decide whether a tool can help answer. A description can be sent even if the assistant does not call that tool. |
| Tool results | Information returned by a tool after it runs. Depending on the tool, a result can include site settings, post content, media details, WooCommerce information, or other data the tool is permitted to read. |
Tool results are treated as site data, not instructions. They can still contain private or personal information, so consider what each enabled tool can access and who can use it.
Examples of how context is added
If someone asks, “What is the status of this draft?”, the provider receives the question and relevant conversation context. If the assistant uses a tool to inspect the draft, the tool’s result is also sent so the provider can explain what it found.
If someone asks for a WooCommerce stock summary, the provider may receive the request, the available WooCommerce tool descriptions, and the inventory details returned by the tool, if that tool is enabled and used. MazeGraph checks WordPress permissions and the workspace’s tool policy before allowing a tool to run.
In Content Studio, a stage may send the brief, relevant instructions, earlier stage material, and any site or tool results needed for that step. Do not put passwords, API keys, private customer details, or other information in a prompt or note unless you are authorized to share it with the connected provider.
Long conversations and older tool results
The Long conversations controls in Settings → Limits and privacy affect which earlier messages and tool results remain in later requests:
- Recent messages only keeps the recent conversation window that fits the context limit.
- Drop old tool results removes older tool output from the context while retaining recent turns. The assistant may need to call a tool again to retrieve older information.
- Summarize older messages replaces older conversation material with a summary. The summary is still context sent to the provider.
These options change future requests. They do not recall information that a provider has already received. See Limits and privacy settings for the full controls.
Credentials and provider handling
Provider credentials are kept by WordPress Connectors. MazeGraph does not read, store, log, or display those keys, and they are not included in the model context. The WordPress AI Client and connector handle the provider call; MazeGraph does not call provider APIs directly. A connector may also contact its provider to list models or check availability.
The provider receives the request content described above and handles it under its own terms and privacy practices. MazeGraph does not send chat data or usage analytics to MazeWP servers and does not collect telemetry. Review the connector and provider policies for their data retention, processing, and billing terms. See Connect an AI provider for connector setup and Data and privacy for MazeGraph’s on-site storage and retention controls.
If a request includes information you did not expect
What happened: A provider’s answer refers to information that was not typed into the latest message. Reason: Relevant conversation history, instructions, or a tool result may have added that information to the request. Solution: Review the conversation, Activity, enabled tools, and long-conversation setting; disable tools the workspace does not need and adjust future context before sending another request. Information already sent cannot be recalled through MazeGraph. Learning: Tell users that tool results and earlier conversation context may go to the connected provider, and enable only tools needed for their work.
For how tool access is governed, read Who may use tools, Approval modes, and Tools.