Dify self-hosted on Windows 11: connect QCode step by step (13 screenshots)

Run Dify on Windows 11 with Docker Desktop and add QCode as an OpenAI-compatible model: 13 numbered screenshots from install to first reply.

Updated 2026-11-23

Need an API key? Sign up →Endpoints & base URLs →

On This Page

Short answer: In Dify, open Integrations → Model Provider, install the OpenAI-API-compatible plugin, then click Add Model and fill in the base URL https://api.qcode.cc/openai/v1, the model name gpt-6.2-sol, your cr_ key, and Chat as the completion mode.

Item Fill in
Integration to install OpenAI-API-compatible
API Base URL https://api.qcode.cc/openai/v1
Model Name gpt-6.2-sol
Model Type / Completion mode LLM / Chat
API key cr_xxxxxxxx

Checked on 2026-11-22 · Dify 1.20.0 · Windows 11

Follow the 13 screenshots below on a Windows 11 PC and you end up with QCode answering inside Dify: a Chatbot app in Studio that answers your first message. Dify itself needs no model-vendor account — the only credential here is a QCode API key.

  • Applies to: Windows 10 / 11 (64-bit) with Docker Desktop and WSL 2; the screenshots show Windows 11
  • Software version: Dify Community Edition 1.20.0 + OpenAI-API-compatible plugin 0.0.87
  • Screenshots updated: 2026-11
  • Time needed: about 25 minutes, plus the first Docker image download
  • You will need: a QCode API key starting with cr_

Before the first screenshot, check that you have these:

  • A Windows 10 or 11 PC, 64-bit, that can open websites, with at least 2 CPU cores and 4 GiB of memory kept for Docker
  • A browser — Edge or Chrome — because Dify has no desktop app; the whole tutorial runs in a browser tab
  • A Dify instance you can sign in to: your own Docker Compose deployment on http://localhost:8080, or Dify Cloud at cloud.dify.ai
  • A QCode API key copied from your dashboard (cr_xxxxxxxx); this is the only item you cannot skip. No account yet? Sign up first, pick a plan on the pricing page, then create the key in your Dashboard.

Download and install Dify

Dify is a web service, so there is no Windows installer for the Dify interface. You either open Dify Cloud in the browser, or you run the same interface locally with Docker Compose. Both routes end on the same Model Provider page used by the steps below. Pick the tab that matches where you are:

Open dify.ai and use the Cloud route: sign up at cloud.dify.ai and the whole tutorial runs in the browser.

For a local deployment, open the official Docker Compose guide at docs.dify.ai, install Docker Desktop from docker.com (Windows, x64), accept the agreement and keep the Use WSL 2 backend. The installer needs administrator rights and usually asks for a restart. The guide's repository also has GitHub Releases, where the Windows file difyctl-v1.20.0-windows-x64.exe is a command-line tool for administrators — it is not the Dify interface, so do not expect a window to open from it. After docker compose up -d, the interface is at http://localhost:8080.

Open dify.ai and use the Cloud route: the interface then lives at cloud.dify.ai, and nothing is installed on the PC.

If you want the interface on your own machine instead, open the official Docker Compose guide at docs.dify.ai and follow it with Docker Desktop. Install Docker Desktop from docker.com first. If a page is slow to load or a download stalls, wait a minute and start it again — the page and the installer are served from the same host, so retrying later usually fixes both.

Read the requirements page first

Before downloading anything, open the official Docker Compose guide. ① the page title confirms you are on the self-hosting page for Docker Compose, and ② the prerequisites block lists what the machine must provide: at least 2 CPU cores, 4 GiB of RAM and Docker Compose itself.

The official Dify Docker Compose self-hosting guide with the page title at the top and the hardware and software prerequisites block below it

Done right you will see: ① the heading of the Docker Compose install page, and ② the prerequisites list with the CPU, memory and Docker Compose lines readable on one screen.

Stuck? The page shows a different install method

Dify documents Docker Compose and Kubernetes side by side, and search results sometimes land on the Kubernetes page. Scroll to the top and check that the address contains docker-compose; that is the route this tutorial uses.

Install Docker Desktop

Dify runs in containers on Windows, so Docker Desktop comes first. On docker.com, click ① Download Docker Desktop, run the downloaded Docker Desktop Installer.exe, accept the agreement and keep the Use WSL 2 backend selected. The installer asks for administrator rights and usually offers a restart; take it.

The official Docker Desktop download page with the Download Docker Desktop button

Done right you will see: the download starts from the ① button, and once installed the whale icon appears in the notification area with the status Engine running.

Stuck? Windows shows “Windows protected your PC”

Click More info, then Run anyway. SmartScreen shows that box for installers it has not seen often enough; the file is the one the ① button just downloaded.

Sign in to your Dify instance

Open http://localhost:8080 in your browser. The very first visit goes to the /install page, where you create the administrator account with an email address and a password; every visit after that opens the sign-in page. Type the email and the password into ①, then click ② Sign in.

The sign-in page of a self-hosted Dify instance with the email and password fields and the Sign in button

Done right you will see: ① both fields filled with dots covering the password, ② Sign in turning blue, and after the click the Dify Home page with the workspace name in the top-left corner.

Stuck? The browser says the site cannot be reached

The containers are not running yet. Open Docker Desktop, wait for Engine running, then run docker compose up -d in the folder of the Dify source and refresh the page. The first start pulls several images and takes a while.

Open Integrations

Dify Home opens after sign-in. The whole navigation sits in one column down the left side: Home, Studio, Agents, Knowledge, Contacts, Skills, Integrations, Marketplace. Click ① Integrations; Dify takes you straight to Model Provider.

Dify Home with the left sidebar navigation and the Integrations entry highlighted

Done right you will see: ① Integrations highlighted in the left column, and the page behind it titled Model Provider with a Search models... box at the top.

Search for the OpenAI-API-compatible provider

Model Provider lists every integration Dify can talk to. Click ① the Search models... box and type OpenAI-API-compatible. ② the matching card appears in the list, with a blue Install button on it.

The Model Provider page with OpenAI-API-compatible typed into the search box and the matching provider card listed below

Done right you will see: ① your text in the search box, and ② one card named OpenAI-API-compatible with an Install button — not OpenAI, which is a different plugin.

Stuck? The search returns nothing

Type only OpenAI and scroll the list. Installing plugins needs the workspace owner or an administrator, so a member account may see the cards without the Install button. After installing, refresh the page before you search again.

Confirm the plugin installation

Clicking Install on the card opens a dialog titled Install Integration instead of installing at once. Read the name in the dialog and click ① Install in the dialog footer. Dify then downloads the plugin from its Marketplace, which takes a few seconds.

The Install Integration confirmation dialog with the blue Install button in its footer

Done right you will see: ① the dialog closes by itself, and the OpenAI-API-compatible card moves into the Configuration required group.

Stuck? The dialog spins and then fails

The plugin is fetched from marketplace.dify.ai. Check that the address opens in another tab; on a company network or behind a proxy the download can be blocked, and retrying on another network usually works.

Open the Add model dialog

The installed plugin now sits under Configuration required with the note Set up api-key or add model to use. Click ① Add Model on the right of that card — this is where the QCode address and the key go in.

The OpenAI-API-compatible card under Configuration required with the Add Model button on its right

Done right you will see: ① Add Model, and clicking it opens a dialog named Add model with Model Name, Model Type and API Key on it.

Name the model and choose its type

Type gpt-6.2-sol into ① Model Name — that is the QCode model ID, so copy it exactly as written. Open ② Model Type and pick LLM. Then click the API Key field and paste your key, which starts with cr_; the dialog hides it behind dots.

The top half of the Add model dialog with gpt-6.2-sol in Model Name, LLM in Model Type and the API Key field filled with dots

Need an API Key?

If you do not have a QCode account yet, Sign up first. If you do not want a monthly plan, buy a pack from the Pay-as-you-go Packs page, then create the key in the Dashboard. Copy the string starting with cr_ and paste it in place of cr_xxxxxxxx. You can pay by credit or debit card (Visa / Mastercard), Apple Pay / Google Pay, or cryptocurrency — see Payment methods.

Done right you will see: ① gpt-6.2-sol spelled exactly like that, ② LLM in the type field, and a row of dots in API Key — no quotation marks around any of them.

Enter the QCode base URL and the completion mode

Scroll down inside the dialog. Put https://api.qcode.cc/openai/v1 into ① API Base URL — the whole path belongs in the field, /openai/v1 included, with no trailing slash and no /chat/completions. Leave ② Completion mode on Chat, then click ③ Add to save and validate the key.

The lower half of the Add model dialog with the QCode base URL in API Base URL, Chat in Completion mode and the Add button

Done right you will see: ① the address ends in /openai/v1, ② Chat is selected, and ③ Add closes the dialog without a red error line.

Stuck? Saving stops with 401 or 404

A 401 means the key was cut short or carries a space or quote — copy it from the Dashboard again and retype the field. A 404 means the address has an extra or missing path; it must be exactly https://api.qcode.cc/openai/v1.

Check that the model is enabled

After Add, the plugin leaves Configuration required. Click Show Models under the card to expand the model list: ① the row gpt-6.2-sol LLM CHAT appears with its switch turned on. Then click ② Default Models in the toolbar at the top right of the page.

The expanded Show Models list with the enabled gpt-6.2-sol LLM CHAT row and the Default Models button at the top right

Done right you will see: ① one model row with gpt-6.2-sol, LLM and CHAT on it and the switch on the right turned on, while ② Default Models is a blue button in the top-right corner.

Set it as the default system model

Default Model Settings opens as a dialog. Its first row is ① System Reasoning Model: open it and choose gpt-6.2-sol from the list, then click ② Save.

The Default Model Settings dialog with gpt-6.2-sol chosen for System Reasoning Model and the Save button

Done right you will see: ① gpt-6.2-sol shown in the System Reasoning Model row, and after clicking ② Save the yellow bar Configure a default system model to run applications disappears from the top of the page.

Stuck? The row says the model list is empty

An empty list means the plugin has no enabled model yet — go back one step and expand Show Models to confirm the switch is on. A model that was disabled in that list is not offered here.

Create a Chatbot app

Open Studio in the left column and click Create from Blank. At first only Workflow and Chatflow are offered; click MORE BASIC APP TYPES to expand the rest and pick ① Chatbot. Type QCode Assistant into ② App name, then click ③ Create.

Studio Create from Blank card with MORE BASIC APP TYPES expanded, the Chatbot type selected, the app name QCode Assistant and the Create button

Done right you will see: ① Chatbot framed in blue after you pick it, ② QCode Assistant in the name field, and ③ Create opening the new application.

Stuck? There is no Chatbot option

Chatbot is folded away by default: click MORE BASIC APP TYPES under the two large cards, and Chatbot, Agent and Text Generator appear. Also check that the page is Studio and not Agents.

Send your first message and check the reply

Creating the app opens the Orchestrate page. ① the model selector at the top already reads gpt-6.2-sol CHAT. Type Hello! Please introduce yourself in one sentence. into ③ the Talk to Bot box at the bottom of Debug & Preview and press Send; ② the answer streams into the panel above it.

The Orchestrate page with gpt-6.2-sol CHAT in the header, the Talk to Bot input box and a streaming reply in Debug and Preview

Done right you will see: ① gpt-6.2-sol CHAT in the header, ③ your sentence in the input box, and ② a short English answer appearing word by word — not a 401, 404 or “model not found” message.

Stuck? The answer never starts

Open Default Model Settings again and confirm System Reasoning Model still holds gpt-6.2-sol; an app created before the default was saved starts without a model. If the reply is red, the key or the address is wrong — see Common errors below.

Verify it works

Your setup is complete when all four of these are true:

  • Model Provider lists OpenAI-API-compatible under the installed integrations, with gpt-6.2-sol enabled in Show Models.
  • Default Model Settings shows gpt-6.2-sol for System Reasoning Model instead of an empty row.
  • QCode Assistant answers in Debug & Preview, and the reply arrives as a stream rather than as a 401, 404 or “model not found” error.
  • Close the browser tab, open Dify again and the model, the default setting and the app are all still there.

All done

You have QCode wired into Dify and a Chatbot app answering through it. From here you can:

Common errors

Error message Why What to do
Set up api-key or add model to use The plugin is installed, but no model has been added to it yet, so Dify has nothing to select Open Model Provider, click Add Model on the OpenAI-API-compatible card and fill in all four fields
Configure a default system model to run applications A model exists but no default was chosen, so new apps start without a model Click Default Models at the top right, set System Reasoning Model to gpt-6.2-sol and click Save
401 The key reached QCode and was rejected: it was copied with a space, a quote or only partly, or it belongs to a plan that has ended Copy the cr_ key from the Dashboard again, paste it into API Key with nothing added, and check the plan state on the Dashboard
404 The address carries an extra or missing path, so QCode receives a request it does not serve Keep API Base URL at exactly https://api.qcode.cc/openai/v1 — no trailing slash, no /chat/completions

FAQ

Do I need a Dify account, or an Anthropic or OpenAI account?

No model-vendor account is needed. Dify Cloud asks for a Dify account; a self-hosted instance asks for the administrator account you create on the /install page. In both cases the key that reaches the model is your QCode cr_ key, and nothing in this tutorial asks for an Anthropic or OpenAI login.

Why is OpenAI-API-compatible missing from the list?

Three usual reasons: the search box still holds a longer phrase (type only OpenAI and scroll), the plugin is already installed and therefore moved out of the Marketplace part of the page, or the account is not a workspace owner or administrator, which is required to install plugins. Refresh the page once before you search again.

Why does the model name have to be gpt-6.2-sol?

Model Name is not a label — Dify sends that string to QCode as the model ID. A QCode model ID must be spelled exactly as the QCode documentation lists it, so gpt-6.2-sol works while GPT 6.2 Sol does not. Any other ID served by the same endpoint goes in the same field, such as gpt-6.2-astra.

Can I use Claude models in Dify through QCode?

Yes, through Dify's own Anthropic integration, with the Anthropic-compatible address https://api.qcode.cc/api instead of the OpenAI-compatible one. The QCode default there is claude-opus-5-5. The 13 screenshots on this page follow the OpenAI-compatible plugin because that is the route covered here end to end; treat the Anthropic address as the option for later.

What happens when my key expires?

It depends on what you bought. A key that came with a personal plan cannot be renewed — reopen the plan, create a new key and paste it into the API Key field again. A key from a pay-as-you-go pack comes back when you buy another pack: the same key keeps working, and nothing in Dify has to be changed.

Does Dify keep the model settings after a restart?

Yes. Models, their keys and the default model are stored in the Dify database, not in the browser session, so signing out or restarting the containers does not remove them. Only deleting the data volumes wipes them, and that also deletes the administrator account and every app.

Should I use Chat or Completion for the completion mode?

Use Chat. QCode serves chat-completions models on this endpoint, and Completion expects a prompt-completion interface that these models do not expose. A wrong choice here shows up as an error on Add rather than in the app.

Models to use

The same QCode key exposes models from several vendors, and the model ID goes into Model Name. Pick one and keep the rest of the fields as they are.

Model Protocol and address Use for
gpt-6.2-sol OpenAI-compatible — https://api.qcode.cc/openai/v1 Default for this tutorial: everyday chat, writing, code questions
gpt-6.2-astra OpenAI-compatible — https://api.qcode.cc/openai/v1 The higher tier of the GPT family, for the hardest tasks
claude-opus-5-5 Anthropic-compatible — https://api.qcode.cc/api Flagship Claude, for long reasoning and large refactors
claude-haiku-5-5 Anthropic-compatible — https://api.qcode.cc/api Lightweight Claude, for fast, cheap answers

A claude-* ID sent to the OpenAI-compatible endpoint is refused, and a gpt-* ID sent to the Anthropic endpoint is refused as well — the model family has to match the address. The live list of IDs is on your Dashboard.

How the address is used

API Base URL is a base, not a full request path. Dify keeps the /openai/v1 part and appends chat/completions itself, which is why the field must stop at /v1.

Field Value Note
API Base URL https://api.qcode.cc/openai/v1 Keep /openai/v1; no trailing slash
What Dify requests https://api.qcode.cc/openai/v1/chat/completions Built from the field above — do not type it
Model Name gpt-6.2-sol Any model ID served on that endpoint
Model Type LLM The type that answers chat requests
Completion mode Chat Not Completion
API Key cr_xxxxxxxx Stored with the model, so it is typed once

Alternative endpoints

If the default host is slow or unreachable from your network, replace only the domain and keep the rest of the address as it is:

Endpoint OpenAI-compatible base Anthropic-compatible base
Default https://api.qcode.cc/openai/v1 https://api.qcode.cc/api
Asia https://asia.qcode.cc/openai/v1 https://asia.qcode.cc/api
North America and Europe https://us.qcode.cc/openai/v1 https://us.qcode.cc/api

Both alternative hosts accept the same key and the same paths. If you switch hosts, edit API Base URL on the model card and click Add again to validate it. Full details are in Endpoints and API paths.

Self-hosting notes

  • Windows needs Docker Desktop plus WSL 2. The Dify repository is a set of services, not an application you can double-click; Docker Desktop runs them, and the guide asks for at least 2 CPU cores and 4 GiB of RAM.
  • The source belongs on the Linux file system. Performance and file permissions behave better when the Dify folder lives inside WSL 2 rather than on C:\.
  • The first start is slow. Several container images are pulled the first time, and the page stays blank until the API container reports ready. Later starts take seconds.
  • http://localhost:8080 is the interface, not the API. Nothing of Dify itself has to be downloaded for the browser; the port is served by the containers.
  • The Dify interface language is separate from the browser. It is switched in the account menu under Preferences → General → Display Language; this tutorial uses the English interface, so every label here matches what the screenshots show.

Get started with QCode

One API key for Claude Code, Codex and more — billed per token, per-model prices on qcode.cc/models

Enterprise →