{"id":170,"date":"2026-10-23T10:58:41","date_gmt":"2026-10-23T10:58:41","guid":{"rendered":"https:\/\/docs.latw.ai\/uncategorized\/mcp\/"},"modified":"2026-10-23T10:58:41","modified_gmt":"2026-10-23T10:58:41","slug":"mcp","status":"publish","type":"post","link":"https:\/\/docs.latw.ai\/pl\/latw-cloud\/mcp\/","title":{"rendered":"Translating from Your AI Agent (MCP)"},"content":{"rendered":"

Available from the Solo<\/strong> plan.<\/em><\/p>\n

LATW Cloud has an MCP server, so the AI agent you already work with \u2014 Claude, ChatGPT, Claude Code,
\nCursor, VS Code, Windsurf, Codex, n8n or your own \u2014 can translate your sites without you opening the
\npanel. Ask it “how much is left to translate into French?” or “translate all missing products into
\nGerman and French”, and the agent checks your site and translates the content itself, with no
\nCredits used (see Translation by the agent itself<\/em>).<\/p>\n

If you would rather have LATW Cloud translate in the background, ask the agent to use LATW’s engine
\n\u2014 with Credits, or with your own API key if the site is set to Your key<\/strong>. The agent also offers
\nit when your plan’s words for agent translation are not enough. It shows you the price first and
\nstarts the job only after you confirm. These are ordinary LATW Cloud jobs: they use the site’s
\nsettings (model, AI source, glossary, prompt) and cost what the same job costs from the panel.<\/p>\n

Either way the jobs show in Jobs<\/strong> with the origin MCP<\/strong> and are written to your live site as
\nthey are translated.<\/p>\n

Connecting your agent<\/h2>\n

There is no key or token to copy. You add LATW Cloud to your agent as a remote MCP server, the
\nagent opens a LATW page, you sign in with your LATW account and press Allow<\/strong>.<\/p>\n

    \n
  • Address: https:\/\/app.latw.ai\/api\/mcp<\/code> (Streamable HTTP)<\/li>\n
  • Sign-in: OAuth \u2014 the agent finds it by itself, nothing else to enter<\/li>\n<\/ul>\n

    The address is on the MCP<\/strong> screen of the panel. Below it, Connect your agent<\/strong> has a tab for
    \nClaude Code<\/strong>, Claude<\/strong>, Codex CLI<\/strong> and ChatGPT<\/strong> \u2014 each with what to copy and where to put
    \nit \u2014 and Setup prompt<\/strong>, a ready-made text you paste into a coding agent such as Cursor: it adds
    \nthe server and starts the sign-in by itself.<\/p>\n

    Claude (Claude.ai, Claude Desktop, mobile)<\/strong> \u2014 open Customize \u2192 Connectors<\/strong>, press + Add<\/strong> and
    \nAdd custom connector<\/strong>. Name it LATW Cloud<\/em>, enter the address, press Continue<\/strong> and keep the
    \nsign-in Claude detects. In a Team or Enterprise organization, an owner adds the connector under
    \nOrganization settings \u2192 Connectors<\/strong> and each member presses Connect<\/strong> next to it. In a chat,
    \nturn LATW Cloud on with + \u2192 Connectors<\/strong>.<\/p>\n

    ChatGPT<\/strong> \u2014 custom MCP servers need Developer mode<\/strong>, which you turn on in ChatGPT’s settings
    \nunder Apps<\/strong> (the plans and labels that offer it are set by OpenAI and change; check ChatGPT’s help
    \nif you do not find it). Then create an app with the address and OAuth as the authentication.<\/p>\n

    Claude Code<\/strong> \u2014 add the server, then run \/mcp<\/code> in Claude Code, pick latw<\/code> and sign in:<\/p>\n

    claude mcp add --transport http latw https:\/\/app.latw.ai\/api\/mcp<\/code><\/pre>\n
    Cursor<\/strong> \u2014 in ~\/.cursor\/mcp.json<\/code> (all projects) or .cursor\/mcp.json<\/code> (one project). Cursor
    \nasks you to sign in when it first connects:<\/p>\n
    {\n  \"mcpServers\": {\n    \"latw\": {\"url\": \"https:\/\/app.latw.ai\/api\/mcp\"}\n  }\n}<\/code><\/pre>\n
    Windsurf<\/strong> \u2014 in ~\/.codeium\/windsurf\/mcp_config.json<\/code>, then press Refresh<\/strong> in Windsurf’s MCP
    \nsettings and sign in:<\/p>\n
    {\n  \"mcpServers\": {\n    \"latw\": {\"serverUrl\": \"https:\/\/app.latw.ai\/api\/mcp\"}\n  }\n}<\/code><\/pre>\n
    VS Code<\/strong> \u2014 in .vscode\/mcp.json<\/code>; VS Code asks you to allow the sign-in when it starts the
    \nserver:<\/p>\n
    {\n  \"servers\": {\n    \"latw\": {\"type\": \"http\", \"url\": \"https:\/\/app.latw.ai\/api\/mcp\"}\n  }\n}<\/code><\/pre>\n
    Codex CLI<\/strong> \u2014 add the server and sign in:<\/p>\n
    codex mcp add latw --url https:\/\/app.latw.ai\/api\/mcp\ncodex mcp login latw<\/code><\/pre>\n
    n8n<\/strong> \u2014 add the MCP Client Tool<\/strong> node to your AI Agent, set the endpoint to
    \nhttps:\/\/app.latw.ai\/api\/mcp<\/code> with the HTTP Streamable transport and choose OAuth2 as the
    \nauthentication. n8n registers itself with LATW (dynamic client registration is on by default in
    \nits MCP credential); connect the credential and allow it.<\/p>\n
    Agents that only run local servers<\/strong> \u2014 bridge with mcp-remote<\/code>, which opens the sign-in in your
    \nbrowser:<\/p>\n
    {\n  \"mcpServers\": {\n    \"latw\": {\n      \"command\": \"npx\",\n      \"args\": [\"-y\", \"mcp-remote\", \"https:\/\/app.latw.ai\/api\/mcp\"]\n    }\n  }\n}<\/code><\/pre>\n
    Your own agent<\/strong> \u2014 use an MCP client library with OAuth support (the official MCP SDKs have it).
    \nThe server publishes where to sign in, and an app registers itself without asking us for keys.<\/p>\n
    Check the connection:<\/strong> ask the agent to call get_account<\/code>. It answers with your plan and your
    \navailable Credits.<\/p>\n
    Allowing an app<\/h2>\n
    When an agent connects, it opens a LATW page: sign in if you are not signed in, check the name of
    \nthe app and press Allow<\/strong>. Cancel<\/strong> sends you back to the app without connecting it. If the page
    \nsays MCP is available from the Solo plan, upgrade first and reload the page. Check that you are
    \nsigned in to the right LATW account \u2014 Switch account<\/strong> is on the same page.<\/p>\n
    What a connected app may do:<\/strong> see your sites, content, jobs and prices; start, cancel, resume and
    \nretry jobs; translate content itself. It uses all sites of your account. It cannot change settings
    \n(model, AI source, glossary, prompt, budget limits, API keys), top up Credits, change your plan or
    \nconnect sites.<\/p>\n
    Connected apps<\/strong> on the MCP<\/strong> screen lists the apps you allowed, with the date. Disconnect<\/strong>
    \ntakes the app’s access away within a minute; jobs it started keep running. To use it again, connect
    \nit again from the app. Jobs started by a connected app show the app’s name in their history, for
    \nexample Job created via MCP, app “Claude”<\/em>.<\/p>\n
    What the agent can do<\/h2>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
    Tool<\/th>\nWhat it does<\/th>\n<\/tr>\n<\/thead>\n
    get_account<\/code><\/td>\nPlan, available Credits, sites used, words left for agent translation, links to the panel<\/td>\n<\/tr>\n
    list_sites<\/code><\/td>\nYour sites with their languages and translation settings<\/td>\n<\/tr>\n
    get_site_coverage<\/code><\/td>\nTranslated, outdated and missing items per language<\/td>\n<\/tr>\n
    list_content_groups<\/code>, list_content<\/code><\/td>\nWhat the site has and the translation status of each item \u2014 the same lists as the Content<\/strong> tab<\/td>\n<\/tr>\n
    list_models<\/code><\/td>\nModels and their price in your plan<\/td>\n<\/tr>\n
    estimate_translation<\/code><\/td>\nCounts the words and prices a job, without spending anything<\/td>\n<\/tr>\n
    start_translation<\/code><\/td>\nStarts the job that was priced<\/td>\n<\/tr>\n
    list_jobs<\/code>, get_job<\/code><\/td>\nProgress, cost and failed items with the reason<\/td>\n<\/tr>\n
    cancel_job<\/code>, resume_job<\/code><\/td>\nAs Cancel<\/strong> and Resume<\/strong> on the job’s page<\/td>\n<\/tr>\n
    estimate_retry<\/code>, retry_failed<\/code><\/td>\nAs Retry failed<\/strong> on the job’s page: first what the retry would cost, then the retry after you confirm<\/td>\n<\/tr>\n
    start_agent_translation<\/code>, get_source_texts<\/code>, submit_translations<\/code><\/td>\nTranslation by the agent itself \u2014 see below<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
    The agent can pick items one by one or ask for everything with a status in a group \u2014 for example,
    \nall products missing in German. One job takes at most 500 items this way; for more, the agent splits
    \nthe work into several jobs. Webflow keeps no translation statuses (see Connecting Webflow<\/em>), so on a
    \nWebflow site the agent picks the items one by one.<\/p>\n
    Price first, then the job<\/h2>\n
    This applies when the agent translates with LATW’s engine. An agent cannot spend Credits in one step. It first asks for an estimate: your site counts the
    \nwords, and the agent gets the cost, your balance after the job and any warning (not enough Credits,
    \na budget limit the job would pass). It must show you that and start the job only after you confirm.<\/p>\n
    The price holds for 30 minutes. If the content grew in the meantime and the job would cost more,
    \nthe job does not start and the agent asks for a new price. As with jobs from the panel, Credits are
    \nreserved when the job starts and what it does not use goes back when it ends; budget limits apply
    \nthe same way.<\/p>\n
    Jobs are translated by LATW Cloud and written to your live site<\/strong> as they finish \u2014 exactly as when
    \nyou start them in the panel.<\/p>\n
    Translation by the agent itself<\/h2>\n
    When you ask the agent to translate, by default it translates with its own model \u2014 the
    \nsubscription you already pay for \u2014 instead of LATW’s engine. LATW Cloud gives it the texts with the site’s prompt, website description and
    \nglossary terms, checks every translation it hands back and writes the valid ones to your site.<\/p>\n
      \n
    • No Credits are used.<\/strong> Instead, the words count towards a monthly limit of your plan, shown on
      \nthe MCP<\/strong> screen and in the plan comparison. When the words left are not enough for a job, the
      \nagent tells you and offers LATW’s engine, with the price first.<\/li>\n
    • It uses your agent’s own model and time.<\/strong> A large job, such as a whole catalog, takes the agent
      \nmany steps and uses up your agent subscription’s limits. For large jobs, LATW’s engine may be the
      \nbetter choice: it translates in the background while you do something else.<\/li>\n
    • What LATW checks:<\/strong> the same keys as the source, HTML tags and their order, shortcodes,
      \nplaceholders, links, empty texts and length limits. A translation that breaks them is not written;
      \nthe agent gets the reason and corrects it. LATW guarantees that the write is technically correct \u2014
      \nthe quality of the translation is your agent’s.<\/li>\n
    • Content that changes<\/strong> while the agent works is read again before its translation is written.<\/li>\n
    • Every platform<\/strong>, WordPress, Drupal, PrestaShop and Webflow. On Webflow the agent works on a few
      \nitems at a time (Webflow limits how many requests a site token may make a minute), translates whole
      \nitems, and what is published follows the site’s Publishing<\/strong> setting (see Connecting Webflow<\/em>) \u2014
      \nas with jobs from the panel.<\/li>\n
    • In Jobs<\/strong> these jobs show the engine Agent (MCP)<\/strong> and no cost. Retry failed<\/strong> is not
      \noffered \u2014 the agent translates the failed items in a new job.<\/li>\n
    • A job the agent stops working on is closed after 24 hours without activity; what it did not hand
      \nin is released on your site.<\/li>\n<\/ul>\n
      The texts are sent to the agent only while it translates and are not stored by LATW Cloud.<\/p>\n
      What the agent cannot do<\/h2>\n
      The agent cannot change settings (model, AI source, glossary, prompt, budget limits, API keys), top
      \nup Credits, change your plan or connect sites. When something needs your hand, it gives you the
      \nlink to the right place in the panel.<\/p>\n
      When something goes wrong<\/h2>\n
      Every error the agent gets has a code and the next step. The common ones:<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
      Code<\/th>\nWhat happened<\/th>\nWhat to do<\/th>\n<\/tr>\n<\/thead>\n
      unauthenticated<\/code><\/td>\nThe agent is not signed in, its sign-in expired, or you disconnected the app<\/td>\nSign in again from the agent (see Connecting your agent<\/em>)<\/td>\n<\/tr>\n
      plan_required<\/code><\/td>\nThe account has no plan that includes MCP<\/td>\nChoose a plan. Your connected apps are kept and work again<\/td>\n<\/tr>\n
      insufficient_credits<\/code><\/td>\nThe job needs more than your available Credits<\/td>\nCredits \u2192 Top up<\/strong> by the missing amount, then estimate again<\/td>\n<\/tr>\n
      not_found<\/code>, job_not_found<\/code><\/td>\nThe site or job does not exist in your account<\/td>\nAsk the agent to list your sites or jobs<\/td>\n<\/tr>\n
      site_unreachable<\/code><\/td>\nYour site did not answer<\/td>\nSee Site Unreachable<\/em><\/td>\n<\/tr>\n
      nothing_to_translate<\/code><\/td>\nEverything chosen is up to date<\/td>\nAsk for “translate again” only if you really want it<\/td>\n<\/tr>\n
      selection_too_large<\/code><\/td>\nMore than 500 items in one job<\/td>\nNarrow it by type or date, or split it<\/td>\n<\/tr>\n
      quote_expired<\/code>, price_changed<\/code><\/td>\nThe price is older than 30 minutes, or the job (or the retry) would now cost more<\/td>\nThe agent asks for a new price<\/td>\n<\/tr>\n
      key_required<\/code><\/td>\nThe site uses Your key<\/strong> without a key chosen<\/td>\nPick a key in Sites \u2192<\/strong> your site \u2192 Settings<\/strong><\/td>\n<\/tr>\n
      word_limit_reached<\/code><\/td>\nYour plan’s words for agent translation are used up this month<\/td>\nUse LATW’s engine, or a bigger plan<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
      If your plan ends, the connected apps stay on the MCP<\/strong> screen, paused, until you renew or
      \ndisconnect them.<\/p>\n","protected":false},"excerpt":{"rendered":"
      Available from the Solo plan. LATW Cloud has an MCP server, so the AI agent you already work with \u2014 Claude, ChatGPT, Claude Code, Cursor, VS\u2026<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_latwm_translated_slug":"","footnotes":""},"categories":[18],"tags":[],"class_list":["post-170","post","type-post","status-publish","format-standard","hentry","category-latw-cloud"],"_links":{"self":[{"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/posts\/170","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/comments?post=170"}],"version-history":[{"count":0,"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/posts\/170\/revisions"}],"wp:attachment":[{"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/media?parent=170"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/categories?post=170"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/docs.latw.ai\/pl\/wp-json\/wp\/v2\/tags?post=170"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}