Skip to content

O'So MCP Server 1.2.0

Choose a tag to compare

@github-actions github-actions released this 03 Nov 19:25
c8c6caa

Added

Reviewing the modules. The new tool somcp_list_modules lists the modules of the shop as Modules > Module Manager shows them, installed or only on disk: name, version, author, description, whether each one is installed, enabled and enabled on mobile, and whether its files are newer than the installed version and wait for an upgrade. It filters like the Status filter and the search field of the screen, and reads only the shop, never the marketplace. It comes with the new Review modules prompt and needs the new permission mcp.modules.read, which the consent screen describes as reading the modules of the shop, and the view right on Modules > Module Manager.

Reading one module. The new tool somcp_get_module reads one module of the shop as its card in Modules > Module Manager shows it - the same version, author, description and state as somcp_list_modules - and adds the hooks it is attached to on the current shop, as Design > Positions lists them. It joins the Review modules prompt and needs the same mcp.modules.read permission and view right on Modules > Module Manager.

Installing a module. The new tool somcp_install_module installs a module whose folder is already on the shop, as the Install button of its card in Modules > Module Manager does, and answers with the module as somcp_get_module reads it afterwards. A module already installed is refused and left as it is, so an install never turns into an upgrade, and nothing is uploaded or downloaded. It joins the Review modules prompt and needs the new permission mcp.modules.write, which the consent screen describes as installing modules already present on the shop, and the add right on Modules > Module Manager.

Enabling or disabling a module. The new tool somcp_set_module_active enables or disables an installed module on the shop, as the Enable and Disable buttons of its card in Modules > Module Manager do, and on PrestaShop 8 switches it on or off for mobile. A request that would change nothing is refused, since PrestaShop would still report a success and run the module's hooks; this module never disables itself. It answers with the module as somcp_get_module reads it afterwards. It joins the Review modules prompt and needs the mcp.modules.write permission, which the consent screen now describes as installing, enabling and disabling modules already present on the shop, and the edit right on Modules > Module Manager.

SEO of the shop's own pages. The new tool somcp_update_meta_page changes the Page title, Meta description and Rewritten URL of one of the pages listed under Shop Parameters > Traffic & SEO - contact, new products, best sales, search and the others - in the languages you approve, in the current shop only, and the Meta keywords on PrestaShop 8, the only version whose form has them. It checks a Rewritten URL as the form does - "Only letters and hyphens are allowed." - instead of reshaping what you approved, requires it in the shop's default language only, accepts the home page's empty one given back unchanged, refuses a friendly URL another page already uses, and warns that a new one makes the old address stop working. It takes the field names somcp_list_meta_pages returns. It joins the Audit and improve SEO prompt and needs the mcp.seo.write permission and the edit right on Shop Parameters > Traffic & SEO.

Reading one content page. The new tool somcp_get_cms_page reads one CMS page as its edit form under Design > Pages shows it: category, title, meta title, meta description, meta keywords (PrestaShop 8), friendly URL, the full page content, Indexation by search engines and Displayed, with the page's address on the storefront. What a page says can now be read and quoted, not only listed, and the SEO fields keep the names somcp_update_cms_seo takes, so a value read can be written back as is. It joins the Read content pages prompt and needs the mcp.content.read permission and the view right on Design > Pages.

Deep thinking, web search and memory, switched per conversation. The + menu of the built-in assistant now has switches for the conversation you are in. Deep thinking asks the assistant to think longer before answering, for Claude, Gemini and OpenAI's reasoning models; it is slower and uses more tokens, and what the assistant thought can be unfolded above its answer. Web search lets Claude and Gemini look things up on the web with the key they already have, and the pages used are listed in the conversation. Claude searches on its own: each search is billed by Anthropic on top of the tokens, and web search must be allowed in your Anthropic console. Gemini searches through a second call to the same model with Google Search, counted in the tokens like any answer. The other assistants - OpenAI, Mistral, DeepSeek, Grok, Cohere - can search too once you pick Claude or Gemini under Web search in the Built-in assistant tab: that assistant searches for them with its key, and its searches count in its tokens. The setting follows the shop levels like the others, and stays greyed until both a Claude or Gemini assistant and an assistant of another provider are set up. With web search on, the assistant can also open a page - one you give it, or one a search found - and read it.

The built-in assistant knows the module it belongs to. Asked how to switch its memory on, it used to search the web and answer about other products, or say it had no memory at all. It now knows it is the module's own chat, what its + menu does and where each setting lives, and points to the module's documentation for the installed version. For super administrators, the employees the settings pages are open to, it can read every setting of the module - each with its tab and label, and whether it is set for this shop, the group or all shops - with the assistants, tools and prompts, and change one setting at a time after you approve it, through the same checks as the settings page. It never reads nor writes an API key. These abilities, like asking another assistant and searching the web, belong to the chat alone: they are not in the module's tool list, and an assistant connected over MCP never sees them. Memory, offered when the memory is switched on in the module settings, can be switched off for one conversation: the assistant then works from that conversation alone, and nothing said in it is added to the memory - what was remembered before stays. A switch only shows when the assistant can use it, and one set away from how a conversation starts shows as a pill under the field, with a cross to undo it. The conversation keeps its switches when you come back to it.

A local assistant, running on your own computer. Add a local assistant in the Built-in assistant tab connects the chat to an AI on your computer - Ollama, LM Studio or any software that speaks the OpenAI API - through O'So AI Relay, a small program you download and run there. Your conversations and your shop data then go to your computer instead of an AI company, with no key to pay for. The relay calls your shop, never the other way round, so nothing has to be opened on your network, and it works whether your shop is hosted online or next to you. The screen opens on a code valid ten minutes, once, with the command to paste in the relay, and a progress bar that moves on by itself: the relay linked, your AI tested, the model to choose. Nothing to click while it waits; Retry shows only when a step gave up, such as a code that expired. Only the models of your AI able to call the shop's tools are offered. The list of assistants shows whether the relay is online, and Test the connection checks it. Removing the local assistant unlinks the relay at once. While nobody uses the chat, the relay asks your shop only every 15 seconds instead of keeping a request open, so it does not hold one of your shop's PHP processes; it is back to answering at once as soon as the chat is opened.

Deleting return statuses. The new tool somcp_delete_return_status deletes one return status the merchant created, as the Delete action of the return status list of Shop Parameters > Order Settings > Statuses does. The five statuses every shop starts with are refused, and so is a status a merchandise return still stands on, with the number of returns. The prompt Change an order status names it too.

Reading the storefront's main menu. The new tool somcp_list_menu_links lists the entries of the main menu as the Selected items of the Main menu module's configuration page show them, in order: categories, products, CMS pages and categories, brands, suppliers, shops and custom links, each with its name in the language asked for and the identifier the other tools take. It also lists the custom links of the Link list, in the menu or not, and names the entries that point to something deleted, which the screen hides. When the Main menu module is not enabled it says so instead of failing. It comes with the new Storefront menu and layout prompt and needs the new permission mcp.appearance.read, which the consent screen describes as reading how the storefront looks, and the view right on Modules > Module Manager.

Adding an entry to the storefront's main menu. The new tool somcp_add_menu_link adds one entry to the main menu, as Add then Save in the Menu Top Link panel of the Main menu module: a category, a product, a CMS page or category, a brand, a supplier, all brands or all suppliers, a shop or a custom link, at the end or at the position asked. A new custom link is created from its label and address in each language and put in the menu in the same call, where the screen leaves it outside the menu. An empty label or address, which the module would accept, is refused, and so is an entry already in the menu. Only the menu of the connected shop changes, and the storefront shows it at once. The Storefront menu and layout prompt names it. It needs the new permission mcp.appearance.write, which the consent screen describes as changing how the storefront looks, and the edit right on Modules > Module Manager.

Changing a custom link of the storefront's main menu. The new tool somcp_update_menu_link changes the label, the address or New window of one custom link of the Link list of the Main menu module, as Edit then Update. Only what is asked changes: where the form empties a field left blank, the other texts and languages are kept, an empty label or address is refused, and a language added to the shop after the link was created gets its text instead of losing it. A link shared by every shop stays shared. A call that would change nothing writes nothing, and the storefront shows the change at once. The Storefront menu and layout prompt names it. It needs the permission mcp.appearance.write, which the consent screen now describes as also changing the custom links of the main menu, and the edit right on Modules > Module Manager.

Reading the storefront's theme and page layouts. The new tool somcp_get_theme reads Design > Theme & Logo for the connected shop: the theme in use with its version and author, the other installed themes, the addresses of the header, mail and invoice logos and of the favicon - saying when the header logo stands in for the mail or invoice logo - and every page of Choose layouts with the layout it uses, and whether that is the theme's default. The Storefront menu and layout prompt names it. It needs the permission mcp.appearance.read, which the consent screen now describes as also reading the theme, logos and page layouts, and the view right on Design > Theme & Logo.

Changing the layout of a page of the storefront. The new tool somcp_update_theme_setting sets the layout of one page - full width, three columns, or two columns with a small left or right column - as one row of Design > Theme & Logo > Choose layouts followed by Save. Only that page changes: every other page keeps its own setting, where the screen saves every row at once. Only the layouts the theme offers are accepted, a call that would change nothing writes nothing, and the storefront shows the new columns at once. It changes the connected shop alone on a multistore installation, and is refused in demo mode. The Storefront menu and layout prompt names it. It needs the permission mcp.appearance.write, which the consent screen now describes as also changing the layout of the pages, and the edit right on Design > Theme & Logo.

Checking the performance settings. The new tool somcp_get_cache_status reads every card of Advanced Parameters > Performance for the connected shop: Smarty, Debug mode, Optional features, CCC, Media servers and Caching, with the caching systems this server can run and the Memcache servers. It warns about what slows the storefront or should not stay on in production: debug mode or the debug profiler on, forced template compilation, the Smarty cache off, overrides disabled, CSS or JavaScript not combined, a caching system the server cannot run. On PrestaShop 9 it says whether debug mode is limited to a debug cookie, without ever giving the cookie's name or value. It changes nothing. It comes with the new Cache maintenance prompt and needs the new permission mcp.system.read, which the consent screen describes as reading how the shop runs, and the view right on Advanced Parameters > Performance.

Clearing the cache. The new tool somcp_clear_cache clears what the Clear cache button of Advanced Parameters > Performance clears: the Symfony cache, the Smarty cache and compiled templates, the XML cache, the combined CSS and JavaScript files of the theme and the class index. The caches are shared by every shop of the installation, so they are cleared for all of them; Memcache or APC is not flushed and no setting changes. The shop is slower for a few seconds while the caches rebuild, so the tool's own description warns of it and keeps a busy shop from being cleared unless the merchant asks. The Cache maintenance prompt names it. It needs the new permission mcp.system.write, which the consent screen describes as clearing the cache, and the delete right on Advanced Parameters > Performance, the same right the button asks.

Changed

Each assistant runs the model you choose, and the chat lets you switch. The Economy, Balanced and Powerful levels are gone: in Built-in assistant, an assistant is now either a Chat assistant with a Default model picked from the list your provider gives, or an Images one. An assistant that already ran keeps the model it was using. In the chat, the assistant button next to + opens a menu like the one of +: every assistant with its default model, a Default badge on the one that answers when none is picked, and for each assistant the list of its models, its default one marked the same way. The model picked there applies to that conversation only.

The sign-in and consent pages can no longer be shown inside another site. These are the pages where a merchant hands the shop to an assistant. PrestaShop lets any site display its pages inside a frame, so another site could have placed the consent button under something else and led a signed-in merchant to click it without seeing it. Browsers now refuse to show these two pages anywhere but on their own.

Claude costs less in a conversation, and the consumption is exact. Assistants on Claude now ask Anthropic to keep the start of each request in its cache: from the second call of a conversation, the tools and the instructions are read back at a tenth of the price or less. The tokens a provider reads back from its cache are counted apart, and the estimate in dollars prices them at that provider's own cache rate (OpenAI, Gemini, DeepSeek, xAI and Mistral cache on their side too), instead of at the full price. The summary the assistant writes of the oldest messages of a long conversation was not counted at all; it now is. Shop Parameters > MCP server > Built-in assistant says how many of this month's tokens were read from a cache.

Brand addresses filtered like their list. somcp_list_manufacturer_addresses now filters like the columns of Catalog > Brands & Suppliers > Brands > Addresses: part of the Last name, First name, Zip/Postal code or City, and one Country by its ISO code. Each address now also gives the name of its country, as the Country column shows it, in the shop default language or the language asked for.

Meta keywords of the shop pages. somcp_list_meta_pages now also gives the Meta keywords of each page on PrestaShop 8; on PrestaShop 9, which no longer has that field, they are shown as empty.

A new look for the built-in assistant. The chat now looks like the AI apps you already know. An empty conversation opens on a short welcome and up to four suggested tasks: the prompts of the page you are on come first, and only those your rights allow are shown. The + button in the field adds files or starts any prompt, and a prompt picked there hands its guide to the assistant with your message. A change waiting for your approval now reads in plain words - each field with its label and its new value, yes or no instead of true or false - with Refuse and Approve side by side and the exact data one click away. The conversations can be searched and are grouped by day, with the month's spending and its budget under them. Icons are drawn as lines instead of emoji, a change still waiting for you shows a clock rather than a check, and an answer can be copied in one click. A prompt declared by another module can name its icon too.

Answers drawn as written. Titles, tables, numbered lists, quotes, separators and blocks of code in an answer of the built-in assistant now show as such, instead of the raw ###, | and --- they were written with. A wide table scrolls sideways inside the chat.

The built-in assistant on a multistore shop. The Built-in assistant tab now follows the shop levels like the other tabs: the assistants, the employees allowed, the monthly cap, the instructions and the memory are set for all shops, a group or a shop, each with the Value of its own here box and "inherited from the group" or "inherited from all shops" beside it. A group or a shop uses the assistants of the level above until it is given its own: checking the box copies them, keys included, to be changed there. A local assistant is not copied: each level links its own relay, so removing a local assistant always unlinks its relay at once. Conversations and spending stay per shop, and a cap set above applies to each shop. With the back office on all shops or a group, the chat now appears as soon as one shop has an assistant, and a new conversation starts by picking its shop on cards showing each shop's logo. The conversation then acts on that shop only, whatever the back office is set to later: its name shows in the chat, each conversation of the list carries its shop number, and the line shown while the assistant works names the shop too.

Assistant spending counted in tokens. The monthly cap of the built-in assistant is now set in tokens, the exact count every AI service reports, instead of an estimate in US dollars that went stale whenever a price changed and counted an unknown model as free. Each assistant can also have a monthly cap of its own, useful to keep a powerful model below an economy one. The chat shows the tokens of the month against the cap, with the estimated cost beside them, marked as an estimate and given to four decimals, since a conversation costs fractions of a cent. A cap set in dollars before is converted on upgrade at the highest price among your assistants, so it is never looser than it was.

The built-in assistant spends a fraction of the tokens. Every message sent the full description of all the shop's tools to the AI service - some 50,000 tokens, a plain hello included, and twice that when a tool was used. The assistant now starts with a tool finder: it looks up the tools a request needs by keyword and loads only those, which the conversation keeps for the messages that follow. A hello now costs a few thousand tokens. A prompt started from the + menu tells the assistant which tools its guide walks through, and each is loaded when the step needs it. Tools switched off on the Tools and prompts tab, or beyond the employee's rights, cannot be found nor loaded. Assistants connected over MCP still see the whole list, as before.

Tools only point to tools that exist. A few tools told the assistant to use a tool not written yet - to list the customer groups, check many links, change an order status, or install and enable modules - so it could reach for a call that only failed. They now say where to look in the back office instead, such as Customers > Groups, and each change still names the tool that shows its result. The link to a new tool is added with that tool.

Shorter order tools. The seven tools of the order page - search, read, change the status, set the tracking number or the private note, add a message, record a payment - now describe themselves in a few sentences instead of a paragraph, about a fifth lighter for the assistant on every step. What matters is kept: when the customer is emailed, which tool gives each identifier, and which read checks a change.

Shorter stock and statistics tools. The three tools of the Stock page - read the stock of a product, list what runs low, change a quantity - and the two statistics tools - sales summary and best-selling products - now describe themselves in a few sentences, a quarter lighter for the assistant on every step. What matters is kept: what each figure means on the screen, which tool gives each identifier, and which read checks a stock change.

Shorter content page tools. The four tools of Design > Pages - list the pages, read one, check their SEO, change their SEO fields - now describe themselves in a few sentences, almost a third lighter for the assistant on every step. What matters is kept: which tool gives each page identifier, the warning that a new friendly URL breaks the old address, and which read checks a change.

Shorter Traffic & SEO tools. The six tools of Shop Parameters > Traffic & SEO - check a public page, read the SEO settings, read and generate robots.txt, list the shop pages and change their SEO - now describe themselves in a few sentences, a third lighter for the assistant on every step. What matters is kept: which tool gives each page identifier, the warning that a new Rewritten URL breaks the old address, the reminder that generating robots.txt loses the lines written by hand, and which read checks a change.

Fixed

A conversation no longer stops for good after the AI service fails. When the AI service failed right after a change you had approved - often a "high demand" error from Gemini - the change was made, but the conversation then refused every new message, with this assistant or any other, and had to be thrown away. The built-in assistant now runs on Neuron AI 4.1.19: the change is recorded the moment it is made, and the next message is answered. A conversation that broke this way before the update picks up again too, without running the change a second time.

An approved change shows as done in the built-in assistant. After the employee approved a change - turning a product on or off, for example - the change was made and the assistant answered, but its step kept the clock of a change still waiting, without "approved" next to it. A refused change showed the same clock. Each step now shows what the employee decided.

An assistant missing a permission is asked to request it, instead of signing in again. When an assistant called a tool its connection was not allowed to use - writing to the catalogue with a read-only connection, for example - the shop answered as if the connection had expired. Some assistants then dropped it and asked the merchant to connect again, which gave the same connection back and the same refusal. The shop now answers that a permission is missing and names it, so the assistant can ask the merchant for exactly that one.

The update to 1.2.0 no longer stops on "Unknown column 'memory_last'". On a shop where an earlier update had not added every column of the assistant's tables, the update failed with that message and the module stayed on its old version. The module now compares its tables with the ones a fresh install creates and adds whatever is missing, on every update. A shop reaches the right tables whatever version it comes from, and an update that was stopped can simply be run again.

The built-in assistant no longer freezes after a slow answer. When the AI service took longer to answer than the shop allows a page to run - often 30 seconds, and on Windows the wait counts - the request was stopped mid-way without a trace: the chat showed "Working..." then nothing, and every new message was refused for ten minutes with "Answer the change waiting for your approval first", though no change was waiting. The chat now asks the shop for the time an answer needs where the host allows it, stops waiting for the AI service before the shop stops the page, and reports a slow answer as an error you can retry. While an earlier step still holds a conversation, the chat says so and waits, and the conversation is freed after five minutes instead of ten.

The test build no longer reports false translation problems. Its summary checked the translations of main's copy too, which the build fetches for its scripts, and showed strings of the last release as missing; it also said "of ." instead of the branch it built. It now checks the module it built, and names the branch.

Settings that depend on each other are checked before they are saved. The memory could be switched on with no embeddings provider, or with Voyage and no key: the page said it was on, and it did nothing. It is now refused with a message saying what is missing; no provider is ever picked for you. A memory that was on and stopped working - its provider's assistant deleted, its key gone - is now flagged in red on the page instead of showing a switch on over nothing. A refresh token set to last less than the access token, which would log every assistant out at its first renewal, and a profile that does not exist are refused too. The rest of the form is still saved, and the message says so. The built-in assistant follows the same rules when it changes a setting, then reads the setting back and tells you whether the feature actually works. It changes only the setting you asked for, and asks you whenever a choice is needed - which provider, which profile - instead of choosing for you.

Gemini's thinking now counts toward the monthly cap. Gemini reports the tokens it spends thinking apart from its answer, and bills them as output; the built-in assistant left them out of the month's count, so a Gemini assistant used more than the chat showed. They are counted now.

Security

The module no longer tells callers which PHP version the shop runs. The answers it sends to a connected assistant - tool calls, sign-in exchanges and the files it hands out - carried the version of PHP behind the shop, which nobody connecting needs. It is now left out of them.


Install: download OSo-MCP-Server-1.2.0.zip and install it from Modules > Upload a module.