Skip to content

EP Assistant

EP Assistant puts an AI chat interface in the PageMotor admin panel and wires it up to nine server-side tools so the customer can manage their entire site in plain English. Write a new page. Check today’s enquiries. Send a follow-up email. Update the brand colours. All through conversation.

Published by ElmsPark Studio.

This is not a chatbot answering support questions. The assistant has real server-side tools that execute actions with safety gates:

  • Asks before it changes anything (1.0.15). Every change to your site’s information, file save, design update or email waits on a Confirm card in the chat, and nothing happens until you press Confirm. The card shows the whole change, however long (1.0.20). Looking things up happens straight away.
  • For site administrators only (1.0.14).
  • Read files within the site root, and save files in your theme folder and your uploads folder (never members-only files), with sandboxing.
  • Run SQL queries against a whitelisted set of tables, with auto-backup before any write.
  • Recompile CSS when the customer asks to change styling.
  • Import theme data the same way EP Provisioning does.
  • Check any page on the site and return its status and content.
  • Send emails via EP Email, rate limited per conversation.

Every tool invocation is logged to an audit table. Every destructive operation takes a backup first.

EP Assistant is designed to be sold through hosting companies as part of an end-to-end AI website product:

  1. The hosting company’s central server runs Discovery AI, which talks to the prospective customer about what they want.
  2. Discovery AI extracts brand voice, audiences, services, and site inventory from the conversation.
  3. The system provisions a small VPS from the hosting company’s catalogue and installs PageMotor + EP Assistant + the rest of EP Suite.
  4. EP Provisioning writes memory files into the new site and configures EP Assistant with an API key.
  5. The customer clicks a welcome-email link, lands in their admin panel, and starts chatting to their site.

The customer never leaves the hosting company’s ecosystem. The hosting company gets to sell a differentiated product.

EP Assistant supports nine providers through three adapter formats:

ProviderFormatWhy pick this
Anthropic (Claude)NativeThe reference provider. Best tool-use quality.
OpenAI (GPT)OpenAIMature API, reliable.
Mistral AIOpenAI-compatEuropean, GDPR-friendly.
DeepSeekOpenAI-compatLowest cost.
GroqOpenAI-compatFastest token throughput.
xAI (Grok)OpenAI-compatAlternative to GPT.
Together AIOpenAI-compatAccess to open-weight models.
Fireworks AIOpenAI-compatFast open-weight serving.
Google GeminiNativeGoogle’s frontier model.

Every provider is configured with API key, model, and max tokens. Swap providers in a dropdown without changing anything else.

  • PageMotor 0.11.5 or later (1.0.20)
  • VPS hosting (not shared). The assistant runs long-polling AJAX and benefits from persistent connections.
  • An API key from one of the nine providers above.
  • EP Email (required for the send_email tool).

Typically installed as part of an automated provisioning flow driven by Discovery AI. If you are installing manually:

  1. Download ep-assistant.zip from the EP Suite downloads page.
  2. Upload via Plugins → Manage Plugins.
  3. Activate.
  4. Open Plugin Settings → EP Assistant and set:
    • Provider. Anthropic Claude is the recommended default.
    • API key. Paste the key for your chosen provider.
    • Model. Leave at the default unless you know what you are doing.
    • Business name. Used in the assistant’s greeting and system prompt.
  5. (Optional) Populate the memory files under src/memory/. If EP Provisioning installed this plugin, these files are already written for you. Otherwise create:
    • brand-voice.md — one page on tone and writing style.
    • audiences.md — who visits your site.
    • services.md — what you sell.
    • site-inventory.md — pages, shortcodes, key content.
    • discovery-context.md — free-form background.

Open the EP Assistant panel from the admin menu. Type into the chat box. The assistant responds with plain English.

Behind the scenes, each message triggers an agentic loop:

  1. Your message + conversation history + system prompt + available tools go to the LLM.
  2. If the LLM wants to call a tool, the plugin executes it server-side and appends the result.
  3. Loop continues until the LLM returns a text-only response.
  4. That response is rendered in the chat with markdown formatting.

The loop has a maximum iteration count so a runaway tool-use cascade cannot spin forever.

ToolPurposeSafety
run_querySQL against the site database.Your own site’s tables only; passwords and keys are hidden from it. A change waits for your Confirm. Auto-backup before any write.
read_fileRead any file inside the site root.realpath() sandboxing. Blocked list: config.php, .env, anything outside ABSPATH.
write_fileWrite or create a file in the theme, uploads or images folders.Waits for your Confirm. Never members-only uploads, never through a link. In uploads and images only safe file types (no web pages, SVG or scripts). No program or server configuration files. Auto-backup of existing content.
recompile_cssTrigger a CSS recompile after theme changes.Localhost-only HTTP call with a shared token.
import_themeImport partial theme data using EP Provisioning’s logic.Key whitelist. Waits for your Confirm.
check_pageHTTP GET any page on your own site.Restricted to your own domain.
send_emailSend an email via EP Email.Waits for your Confirm. Rate limited to 10 per conversation.

Every successful tool call creates an audit row with input, output, and backup data (if applicable).

Type a slash command to load a skill prompt:

CommandPurpose
/update-pageEdit the content of an existing page.
/add-pageCreate a new page with a given slug and content.
/write-articleDraft a blog article.
/write-newsletterCompose an EP Newsletter campaign.
/check-enquiriesReview today’s EP Email contact submissions.
/check-bookingsReview today’s EP Booking appointments.
/check-siteRun the autonomous site-doctor agent — page reachability, broken links, spam signals.
/send-emailCompose and send a one-off email.
/update-stylesChange theme colours, fonts, spacing.
/update-themeBroader theme changes.
/helpShow every available command.

Slash commands load a specific skill prompt that primes the assistant for that task.

The prompt sent to the LLM is built from five layers:

  1. Core rules: no jargon, backup before modify, plain text over corporate language. The English variant the assistant uses is configurable per-install under AI Connection → Assistant → English variant (Auto, British, American, International). Default is Auto, which instructs the assistant to match the client’s existing site variant rather than force one.
  2. Operations guide (hardcoded): How to use each tool correctly.
  3. Plugin knowledge (from EP Support): Full knowledge base for all 18 EP Suite plugins.
  4. Memory files (per-site): The five markdown files described above.
  5. Skill instructions (per-turn): Loaded when the user invokes a slash command.

This layered prompt is why the assistant feels tailored to the customer’s site rather than generic. The memory files carry the business context; the plugin knowledge carries the system context.

  • {prefix}ep_assistant_conversations — conversation history with messages JSON, status, timestamps.
  • {prefix}ep_assistant_audit — every tool call logged with input, output, backup data.

Check the key matches the provider selected in the dropdown. Anthropic keys start sk-ant-, OpenAI keys start sk-, Gemini keys are alphanumeric with no prefix. Paste the full key including the prefix.

LLM provider is slow or down. Check your provider’s status page. Try switching to a different provider temporarily.

The run_query tool has a whitelist of 10 allowed tables. Custom plugin tables are not on it by default. If you need to give the assistant access to more tables, edit the whitelist in class-ep-assistant-tools.php carefully.

“The assistant says it can’t save a file there”

Section titled ““The assistant says it can’t save a file there””

The write_file tool writes only inside user-content/themes, user-content/uploads (1.0.21; never uploads/protected, where members-only files live) and, where the folder exists, user-content/images. In uploads and images it saves only file types no browser runs: pictures other than SVG, PDF, text, CSV, JSON, audio, video, office documents, fonts and zip files. For a web page, an SVG or a script, ask it to save into your theme’s folder. Writing to plugin code or PM core is deliberately blocked.

Every tool call creates a row. Over a busy month this can be thousands of rows. Truncate {prefix}ep_assistant_audit manually if it grows unmanageable.

“I changed the memory files but the assistant doesn’t seem to know”

Section titled ““I changed the memory files but the assistant doesn’t seem to know””

Memory files are read once per conversation at start. Open a fresh conversation to pick up changes. Existing conversations keep their original system prompt for consistency.

For a quick question about this plugin, EP Support inside your admin is the fastest option. The chat widget sits on every EP plugin settings page and knows which one you’re on, with starter questions and links preloaded for that exact screen.

For anything bigger — a bug report, a feature request, or a “how do I…” that needs a real reply — open a ticket at help.elmspark.com. A real person, helped by AI, writes the reply. Usually within a few hours. Tickets don’t disappear into the void.

  • Changing this plugin’s settings over the API now needs an admin token. Producer tokens and AI apps connected through a producer account can still read the settings. Saving from the Admin screen works exactly as before.
  • The current Claude models. The model menu now offers Claude Haiku 5.5, Sonnet 5.5 (recommended), Opus 5.5 and Fable 5.2. A site that chose Sonnet 4.6 or Haiku 4.5 now runs Sonnet 5.5 or Haiku 5.5, the models that replace them.
  • The model list now updates itself. The assistant reads ElmsPark’s central list of AI models once a day, so new and retired models reach your site without waiting for a plugin update. A model you typed in yourself is left exactly as it is.
  • Fixed: after recompiling your site’s CSS, the assistant stopped with “Something went wrong (HTTP 400)” instead of carrying on. It now finishes the job.
  • Fixed: with Mistral, DeepSeek, Groq, xAI, Together AI or Fireworks and no model chosen, the assistant asked for an OpenAI model name. It now uses that provider’s recommended model.
  • Nothing to do after updating. If you want to stay on an older model, it stays selectable on the settings screen, marked with what it now runs as.
  • The assistant can save files in your uploads folder. It could only save in your theme folder and an “images” folder that most newer sites do not have, so asking it to save a file often failed. It can now save in user-content/uploads, where PageMotor keeps your files, for example a price list or a document for visitors to download.
  • Members-only files stay out of reach. It never saves anything in your protected uploads, where members-only files live.
  • Only safe file types are saved. The assistant refuses file types that are not suitable for uploads.
  • You are only asked to confirm saves that can happen. A save to a folder the assistant may not use is refused before a Confirm card appears, with a note saying where it can save instead.
  • Nothing to do after updating.
  • If you are updating from 1.0.13, this update also brings everything in 1.0.14 to 1.0.19 below, including several security fixes: the assistant is for site administrators only, it asks you to confirm before it changes anything or sends an email, and it works only with your own website’s information. Please update promptly.
  • A Confirm card now always shows the whole change. Before, a long file, email or design change was cut short on the card, so you could approve something whose end you had not seen. Now every character is in the card’s scrolling panel, and a long change says how long it is, so you know to scroll to the end.
  • A change too big to show is not offered at all. If a change would be longer than 100,000 characters, the assistant is told to split it into smaller changes, each with its own card.
  • Nothing to do after updating.
  • The change history keeps the whole answer to a large lookup. Before the assistant changes a list of records it looks them up first, as a backup. The history now keeps every row of that lookup (up to about 4 MB), not just the first 100, so you have a full copy to restore from. The assistant itself still reads at most 100 rows at a time, so long lists cannot slow it down or run up your AI bill.
  • Large lookups are no longer missing from the history. A lookup with a big answer used to leave no record at all. Now it is always recorded, and if an answer is too large to keep in full, the history says how big it was.
  • Text saved in an older character format no longer comes back empty. The assistant and the history now see it, with any unreadable letter shown as a placeholder.
  • Nothing to do after updating.
  • Messages from the public are clearly flagged. If a conversation has read text written by visitors, every Confirm card in it shows a yellow warning, so you can check that a change was not asked for by a visitor.
  • Fewer “please try again” errors when the AI service is busy. The assistant now waits as long as the AI service asks before trying again, also when the service gives a time of day rather than a number of seconds, and also when a reply has not started to appear yet. If the service needs longer than a minute, you are told roughly how long to wait.
  • A proper summary when a request takes many steps. When the assistant reaches its step limit or runs short of time, it now tells you what it did, what changed on your website and what is left to do, instead of a stock apology.
  • Large lists no longer overload the assistant. It reads at most 100 rows of a list at a time and is told when there are more, so very large tables do not slow it down or confuse it.
  • A conversation that grows too long now says so and suggests starting a new one, instead of a general error.
  • Nothing to do after updating.
  • A limit on how many messages each admin can send an hour (60 unless you change it under “Messages per hour”), so a stuck browser tab or a shared login cannot run up a large AI bill.
  • Your Google Gemini key is no longer part of the web address the assistant calls, so it stays out of server and proxy logs.
  • The change history is tidied automatically. The record of what the assistant looked up and changed is deleted after 90 days (adjustable under “Keep the change history for”).
  • More private files are off limits. The assistant will not open hidden files and folders, backup copies of your configuration, keys, certificates or database exports.
  • Nothing to do after updating.
  • Security fix: the assistant only works with your own website’s information. It can no longer look into other databases on the same server, or into parts of the database it was not meant to see.
  • Passwords, keys and tokens stay hidden. When the assistant looks up accounts or settings, those values are blanked out before it sees them.
  • Page checks stay on your website. The assistant can check that your own pages load, and nothing else, even when a page redirects.
  • Files stay where they belong. The assistant reads and saves files only inside your website’s folders, and saves only in your images and theme folders.
  • Nothing to do after updating.
  • Security fix: the assistant now asks before it changes anything. When it wants to change information on your website, save a file, update your design or send an email, it shows you a card in the chat saying exactly what will happen, with Confirm and Cancel buttons. Nothing changes until you press Confirm. Looking things up still happens straight away.
  • A change you have not confirmed lapses after 30 minutes. Just ask again if you still want it.
  • The assistant can no longer change user accounts or save program files on your website. Use Admin > Users for accounts.
  • Links in the assistant’s replies now open only ordinary web addresses.
  • Nothing to do after updating.
  • Security fix: the assistant is now for site administrators only. It answers administrators and nobody else, and its chat panel is shown only to them. Please update promptly.
  • The “Enable Assistant” switch is now respected everywhere. With it off, the assistant does not answer at all, even if an API key is saved.
  • Nothing to do after updating.
  • Your API key is now stored encrypted by EP Assistant itself. On PageMotor 0.11.3 and later, core encrypts password fields when you save them, but a password field placed inside a settings section was encrypted in a way core never undoes, so EP Assistant would have read scrambled text instead of your key the first time you saved it. EP Assistant now keeps the key encrypted with the same protection as the rest of the EP Suite, and never shows it back through the admin API.
  • Nothing to do after updating. A key you saved earlier is moved across automatically the first time the site loads, including one core had already scrambled. If the key cannot be recovered, the error log says so and asks you to enter it again.
  • Fixes a doubled cache-busting tag on its chat script and admin script. On PageMotor 0.11 and later the page asked for each file with two version tags on the end of its address (?v=…?v=…), because the core had started adding its own tag and the plugin was still adding one too. The files loaded correctly, so nothing was visibly broken; the addresses were just malformed and every cache saw a longer key than it needed. The plugin now leaves the tag to the core on 0.11 and later, and still adds its own on older cores, where nothing else would.
  • No change to what these files do.
  • Fixes “Your session has expired. Please reload to ensure your security.” on PageMotor 0.11. The message appeared on this plugin’s admin screens even though you were signed in perfectly normally, and whatever you were doing failed to save.
  • Nothing was wrong with your session. PageMotor 0.11 started handling part of the security check that this plugin was already handling itself, and the two together made every save look invalid. The plugin now checks whether PageMotor has already done it.
  • Visitors who were not signed in were never affected, on any version.
  • There is nothing to reconfigure, and nothing else changed.