Skip to main content
Claude Desktop can use Eden AI in two ways, separately or together:
  • Models: Claude Desktop sends its model requests to Eden AI’s Anthropic-compatible endpoint instead of Anthropic, and they are billed to your Eden AI account.
  • Tools: Eden AI’s MCP server gives Claude OCR, translation, speech, web search and the rest of the expert model catalog as tools.
Using Eden AI for the models switches Claude Desktop to what Anthropic calls third-party mode. In that mode:
  • Chat, Cowork and Code all run on the models you configure through Eden AI.
  • The app does not use your claude.ai account: your claude.ai conversations are not shown, and new conversations are stored on your computer.
  • To go back to your claude.ai account, choose the Anthropic sign-in option on the sign-in screen, as described in Anthropic’s setup guide.
Claude Desktop does not take its base URL or key from the Claude Code CLI settings, so configure each one you use.

Prerequisites

Use Eden AI for the models

Claude Desktop accepts only Claude models through a gateway: it rejects other model IDs when you save the configuration. To use GPT, Gemini, DeepSeek or other models as a coding agent, use the Claude Code CLI, which sends whatever model ID you set, or another coding agent such as OpenCode. Note that Anthropic does not support non-Claude models behind a gateway.

1. Open the configuration window

Launch Claude Desktop. You do not need to sign in, and if you are already signed in you do not need to sign out. From the menu bar on macOS, or the ☰ menu at the top left on Windows, choose Help → Troubleshooting → Enable Developer Mode. Then choose Developer → Configure Third-Party Inference.

2. Configure the gateway

Fill in:

3. Add the models

Under Models, add one entry per model: The model ID is what Claude Desktop sends to Eden AI, so the first three need the anthropic/ prefix. The second column tells Claude Desktop which Claude model the ID is, so it uses that model’s context window and effort controls. Without these entries, Claude Desktop tries to build the model list from the gateway, which does not work with Eden AI yet. The last entry is for auto mode in the Code tab, and it is the exception: enter it exactly as claude-sonnet-5, without the prefix. Auto mode’s safety checks call that ID, and Anthropic’s configuration reference says that if the ID is missing from the list, auto mode can block actions it would otherwise allow. Eden AI accepts claude-sonnet-5 as is and routes it to a Claude Sonnet 5 provider.

4. Apply and restart

To add Eden AI’s tools in the same restart, fill in the Connectors section first. Click Apply Changes, then Save & Restart. If the app relaunches on the sign-in screen, choose the option to start with the third-party configuration.

5. Start a session

Open the Code tab, choose a folder for the session to work in, and send a message. The requests appear under Monitoring on app.edenai.run. To change the configuration later, open the same window again. If the Developer menu is gone after the restart, choose Help → Troubleshooting → Enable Developer Mode again.

Use Eden AI’s expert models as tools

Eden AI’s MCP server needs your API key in an Authorization header on every tool call. How you give it that header depends on the mode Claude Desktop runs in.

In third-party mode

If you set up Eden AI for the models above, add the MCP server in the same configuration window:
  1. In the configuration window (Developer → Configure Third-Party Inference), go to Connectors, and under Managed MCP servers choose Add → Blank.
  2. Fill in:
  3. Click Test this connection. It should list the Eden AI tools. Claude Desktop warns that the Authorization header looks like a credential; the warning does not block saving. On a shared or managed machine, use a headers helper instead: a command that prints the headers as a JSON object when Claude Desktop needs them.
  4. Click Apply Changes, then Save & Restart.
The tools are then available in Chat and Code, and in Cowork where it is installed. Ask for something an expert model does, for example “Use the Eden AI translation tool to translate ‘good morning’ to French”.

Signed in with a claude.ai account

A custom connector (Add custom connector) does not work with Eden AI: the form has no field for an API key, so the tools list but every call fails with “No Eden AI API key was provided”. Add the server to Claude Desktop’s configuration file instead, through the mcp-remote bridge, which passes the key as a header. This needs Node.js installed.
  1. Open Settings → Developer → Edit Config. This opens claude_desktop_config.json. Add an edenai entry under mcpServers, keeping whatever else the file already holds:
    Write Authorization:${AUTH_HEADER} without a space after the colon. Claude Desktop on Windows does not escape spaces inside args, which breaks the header. On Windows, this PowerShell snippet adds the entry for you to %APPDATA%\Claude\claude_desktop_config.json, the file Edit Config opens. It saves a copy of the file first, asks for the key so it never lands in your shell history, and keeps the rest of the file. If Edit Config opens a file at a different path, put that path in $cfg:
  2. Quit Claude Desktop completely and start it again: it reads the file only at startup, and closing the window leaves it running. On macOS, press Cmd+Q. On Windows, click the ^ arrow next to the clock, right-click the Claude icon and choose Exit.
  3. In a new chat, open the + menu and switch edenai on, then ask for a translation as above.

Troubleshooting

”Couldn’t sign in to gateway” or “the provider rejected your credentials”

Claude Desktop showed this for claude-sonnet-5-5 entered without the anthropic/ prefix, and a wrong or revoked key also gives a credentials error. Check the model IDs first, then test the key outside Claude Desktop:
OK means the key and the model work, so the problem is in the configuration window: re-enter the key and check the model IDs. A 401 means the key is wrong. A 402 means the account is out of credits or over a spending budget; Anthropic’s gateway guide says Claude Desktop shows a Billing issue card for it. A 403 means a policy on your Eden AI account blocks the model, and a 451 that the model is not available in the region your account is restricted to.

”Model(s) not found or inactive”

The model ID that reached Eden AI is not in the catalog. Check that it matches the table in step 3 exactly: anthropic/claude-sonnet-5-5 works, while a bare claude-sonnet-5-5 or a dot in the version does not. The claude-sonnet-5 entry for auto mode is the one that stays without a prefix. If an ID looks right, delete the entry and type it again.

”Couldn’t load your organization model list”

This appears when the configuration has no model entries. Add the entries from step 3.

”Reinstall the desktop app to access Cowork”

On Windows, Cowork needs the .msix package of Claude Desktop; see Anthropic’s installation guide. The Code tab works without it.

The Eden AI tools do not appear

In third-party mode, open the Connectors section and click Test this connection to see the error. Connection events are logged in mcp.log, in ~/Library/Logs/Claude-3p/ on macOS and %LOCALAPPDATA%\Claude-3p\logs\ on Windows. Signed in with a claude.ai account, Claude Desktop writes mcp-server-edenai.log in ~/Library/Logs/Claude/ on macOS. On Windows, look in %LOCALAPPDATA%\Claude\logs\, or in %APPDATA%\Claude\logs\ on some installs. If there is no such log, look for errors about edenai in mcp.log in the same folder, check that claude_desktop_config.json is valid JSON, and make sure you quit Claude Desktop fully before starting it again. If mcp-server-edenai.log shows npx not found, install Node.js and restart Claude Desktop.

Next Steps

  • Claude Code: the Claude Code CLI on Eden AI, with any model in the catalog
  • MCP server: the full tool catalog and how its tools work
  • Prompt Caching: how caching works on the Messages endpoint