Claude Desktop + Ollama Cloud

Run Ollama Cloud models inside Claude Desktop's embedded Claude Code sessions on macOS.

Claude Desktop does not currently expose a normal setting to swap its built-in Claude model for an Ollama Cloud model. This workaround routes Claude Desktop's embedded Claude Code / local code sessions through Ollama's Anthropic-compatible local gateway.

It turns this:

Claude Desktop local code session -> Claude Code -> Anthropic default model

into this:

Claude Desktop local code session -> wrapper -> Claude Code -> Ollama -> deepseek-v4-pro:cloud

The 30 Second Version

ollama signin
ollama list | grep ':cloud'

security add-generic-password \
  -a "$USER" \
  -s "ollama-cloud-api-key" \
  -w "PASTE_YOUR_OLLAMA_API_KEY_HERE" \
  -U

./scripts/install.sh
./scripts/smoke-test.sh

Then open Claude Desktop, start a Claude Code / local code session, and watch:

tail -f /tmp/claude-desktop-ollama-wrapper.log

If the log says routed Claude Code launch to ...:cloud, the route is working.

Important Scope

This does not replace the normal Claude chat model in Claude Desktop.

It targets the Claude Code / local agent sessions that Claude Desktop launches internally. Those sessions can edit files, run tools, and work in local code folders.

Tested On

• macOS on Apple Silicon
• Claude Desktop 1.6608.0
• Claude Code CLI 2.1.140
• Ollama 0.21.2
• Ollama Cloud model deepseek-v4-pro:cloud

This is an unsupported workaround. Claude Desktop and Ollama change quickly, and a Claude Desktop update may overwrite the helper or change the launch path.

Share Your Setup

If this works for you, please open a tested configuration issue with your macOS, Claude Desktop, Claude Code, Ollama, and model versions.

That helps other users see which combinations are working.

Why This Is Useful

• Try Ollama Cloud models inside a familiar Claude Desktop code workflow.
• Use models like deepseek-v4-pro:cloud, kimi-k2.6:cloud, glm-5.1:cloud, or qwen3.5:397b-cloud.
• Keep the API key out of plaintext files by storing it in macOS Keychain.
• Restore normal Claude Desktop behavior with one command.

How It Works

Claude Desktop launches a helper named:

/Applications/Claude.app/Contents/Helpers/disclaimer

The installer preserves the original helper as:

/Applications/Claude.app/Contents/Helpers/disclaimer.real

Then it installs a small Bash wrapper at the original path. When Claude Desktop starts an embedded Claude Code session, the wrapper:

1. Reads your Ollama API key from macOS Keychain.
2. Sets ANTHROPIC_BASE_URL=http://127.0.0.1:11434.
3. Rewrites Claude model arguments such as claude-sonnet-4-6 to your selected Ollama Cloud model.
4. Launches the host Claude Code CLI.
5. Logs the route to /tmp/claude-desktop-ollama-wrapper.log.

flowchart LR
  A["Claude Desktop"] --> B["helper wrapper"]
  B --> C["Claude Code CLI"]
  C --> D["Ollama local gateway"]
  D --> E["Ollama Cloud model"]

Loading

Prerequisites

Install Ollama and sign in:

ollama signin
ollama --version

Confirm cloud models are available:

ollama list | grep ':cloud'

Confirm Claude Code is installed:

"$HOME/.local/bin/claude" --version

If needed:

curl -fsSL https://claude.ai/install.sh | bash

Save your Ollama API key in Keychain:

security add-generic-password \
  -a "$USER" \
  -s "ollama-cloud-api-key" \
  -w "PASTE_YOUR_OLLAMA_API_KEY_HERE" \
  -U

Do not commit or paste your real API key.

Install

./scripts/install.sh

The installer defaults to the first Claude model listed in:

~/.ollama/config.json

If no model is found there, it falls back to:

deepseek-v4-pro:cloud

Override the model at install or launch time:

OLLAMA_CLAUDE_MODEL="kimi-k2.6:cloud" ./scripts/install.sh

Test

Run the terminal smoke test:

./scripts/smoke-test.sh

Expected output:

ok

Then open Claude Desktop and start a Claude Code / local code session. Ask:

What model are you running? Then create /tmp/ollama_desktop_test.txt with "hello from ollama desktop" and read it back.

Verify:

cat /tmp/ollama_desktop_test.txt
tail -3 /tmp/claude-desktop-ollama-wrapper.log

The wrapper log is the reliable proof. A model may not always self-report accurately.

Switch Models

Edit the Claude integration model in Ollama's config:

jq '.integrations.claude.models = ["glm-5.1:cloud"]' \
  "$HOME/.ollama/config.json" > /tmp/ollama-config.json \
  && mv /tmp/ollama-config.json "$HOME/.ollama/config.json"

Or override per launch:

OLLAMA_CLAUDE_MODEL="qwen3.5:397b-cloud" open -a Claude

Restore

./scripts/restore.sh

This moves disclaimer.real back to disclaimer and restores normal Claude Desktop behavior.

Troubleshooting

No wrapper log:

tail -f /tmp/claude-desktop-ollama-wrapper.log

Then start a Claude Desktop local code session.

No cloud models:

ollama signin
ollama list | grep ':cloud'

Authentication fails:

security find-generic-password -a "$USER" -s "ollama-cloud-api-key" >/dev/null && echo ok

Claude Desktop update broke it:

./scripts/install.sh

Need to undo everything:

./scripts/restore.sh

Security

• Store the Ollama API key in Keychain.
• Do not paste the key into chat, shell history, screenshots, or GitHub.
• Rotate the key if it was exposed anywhere.
• This modifies the Claude Desktop app bundle. Keep the restore script nearby.

Related

The original tutorial gist is here:

https://gist.github.com/FernellysM/2ce62ea017ae9d78d7b2c54603e718c6

License

MIT