HopBase DocsClaude Desktop
Connect Chat, Cowork, and Code in Claude Desktop to HopBase through its third-party inference mode.
Claude Desktop is Anthropic's desktop app. Its third-party inference mode (Claude Desktop on 3P) can send requests to any gateway that implements the Anthropic Messages API, and HopBase connects as that gateway.
| Setting | Value |
|---|---|
| Inference provider | Gateway |
| Gateway base URL | https://api.hop-base.com (no /v1) |
| Gateway API key | A key from a Claude group |
| Credential kind | Static API key (default) |
| Gateway auth scheme | Bearer (default) |
Before you start
Install Claude Desktop from claude.com/download. The official requirements are macOS 14 or later, or Windows 10 version 2004 or later. On Windows, use the .msix installer; the older .exe installer comes without Cowork.
On the console's API Keys page, prepare a key from the Claude Max (Official, Full Quota) group. The Claude Max (ccmax) group only admits the Claude Code client, so do not use it in Claude Desktop.
Third-party inference mode does not sign in to an Anthropic account, and conversations stay on your computer. claude.ai on the web, the mobile apps, and voice mode are not available in this mode.
Setup
Turn on Developer Mode
Launch the app and stay on the sign-in screen without signing in. From the macOS menu bar, or the ☰ menu at the top left of the Windows sign-in screen, choose Help → Troubleshooting → Enable Developer Mode.
Open the third-party inference window
Choose Developer → Configure Third-Party Inference….
Fill in the connection
Under Connection, set Inference provider to Gateway, then enter the base URL and key from the table above. Leave the other two fields at their defaults.
Apply and restart
Click Apply Changes, then Save & Restart. After the restart, choose to start in third-party inference mode on the sign-in screen.
Send a test message
Start a new conversation and ask for a one-word reply. When it succeeds, the request shows up in the console's Usage page.
Choose models
You can leave the model list empty: at launch the app reads GET /v1/models to fill it. If the picker is empty or lists models that fail, enter model IDs by hand under Models in the configuration window. The first one is the default model.
claude-opus-5
claude-sonnet-5Only enter IDs that GET /v1/models returns for your key. IDs are case-sensitive.
Which features work
According to the official feature table, Chat, Cowork, and Code all call models through the gateway in third-party inference mode, so all of them run on HopBase. Projects, Artifacts, local and remote MCP, and skills and plugins also work.
More details
Sign out in the app to get back to the sign-in screen, then follow the steps above. To return to the regular mode, choose the Anthropic sign-in on the sign-in screen; your third-party inference configuration is kept.
Per the official docs, when a gateway does not forward Anthropic server tools, web search needs a separate search provider set up in the configuration window first. Skip this if you do not need web search.
- The claude.ai sign-in screen still appears after the restart: the configuration was not read. Quit the app completely, reopen it, and check that Inference provider is Gateway.
- To see what the app actually read, choose Help → Troubleshooting → Generate Diagnostic Report, export it, and check
managed-config.txtandprovider-status.txt. - If your company has deployed a managed Claude configuration to the computer, the window is read-only. Contact your administrator.
- 401: the key is wrong, or it does not belong to a Claude group.
- 404: the base URL includes
/v1, so requests go to/v1/v1/messages.
For other errors, see Troubleshooting.
Next steps
- Claude Code: use the same Claude models in a terminal
- Anthropic SDK: call Claude from your own code
- Troubleshooting: look up an error by its message