ollama-scripts

Point coding CLIs at ollama.com's cloud models using one OLLAMA_API_KEY. No ollama launch, no sign-in.

It is however recommended that you have the tool curl -fsSL https://ollama.com/install.sh | sh installed so you can take advantage of local models.

Setup

export OLLAMA_API_KEY=...   # your ollama.com key (put in shell rc)
# set -gx OLLA...           # if you're on fish
mise install                # fetch gum (model chooser)
mise run check              # verify key + API reachable
mise run install            # symlink launchers into ~/.local/bin

mise run uninstall removes the symlinks. No mise? brew install gum and ln -s "$PWD"/ollama-{claude,codex,pi,hermes,oh-my-cli,oh-my-cli-app,codex-app,claude-app,gui} ~/.local/bin/.

Launchers

Command	Harness	Endpoint
ollama-claude	Claude Code	Anthropic /v1/messages (Bearer)
ollama-codex	Codex CLI	OpenAI /v1/responses
ollama-pi	pi	OpenAI /v1/chat/completions
ollama-hermes	Hermes agent TUI	OpenAI /v1/chat/completions (ollama-cloud)
ollama-oh-my-cli	oh-my-cli	OpenAI /v1/responses
ollama-oh-my-cli-app	oh-my-cli Desktop	Electron shell, env-pumped to ollama.com

Pick a model three ways (first wins): --model NAME, OLLAMA_MODEL=NAME, or the gum chooser (prefilled with your last pick, remembered per-harness in ~/.config/ollama-scripts/). Everything else is passed through:

ollama-codex --model qwen3-coder:480b exec "fix the failing test"
ollama-claude                       # chooser, then normal claude session

Desktop apps

Same idea for the GUIs, mirroring ollama launch but keyed straight to ollama.com — no sign-in. macOS only.

Unlike the CLIs, desktop apps can't read the shell env, so these must write config files — persistent, not ephemeral. Every touched file is copied to <file>.ollama-scripts.bak first, and ollama-unset.sh reverts everything.

Command	App	How
ollama-codex-app	Codex desktop	writes ~/.codex/config.toml provider (backed up), then restarts Codex
ollama-claude-app	Claude Desktop	writes Claude's 3p gateway profile (backed up), then relaunches
ollama-gui	Ollama app	launches Ollama.app with OLLAMA_API_KEY in its env

ollama-codex-app --model glm-5.2        # picks a model like the CLIs
ollama-claude-app                       # switch Claude Desktop to Ollama Cloud
ollama-gui                              # open the Ollama app with cloud access

Because the Codex launchers write a model catalogue covering all your ollama cloud models, you can switch mid-session with /model inside Codex instead of restarting the harness. Launching still picks one default (--model, OLLAMA_MODEL, or the chooser), but the picker lists everything.

Undo

ollama-unset.sh returns Codex, Claude Desktop, and pi to their original providers — restoring each config from its .ollama-scripts.bak (or stripping only what was added if no backup exists) and quitting the apps so they reload clean. The CLI wrappers (ollama-claude, ollama-codex) need no undo; they only set env for their own subprocess.

ollama-unset.sh                         # revert all app/pi config changes

Per-app reverts also exist: ollama-codex-app --restore, ollama-claude-app --restore.

VS Code

The Ollama VS Code extension supports cloud models natively — no launcher script needed. Install the extension, then set two options in VS Code settings (Cmd+,):

• ollama.endpoint → https://ollama.com
• ollama.headers → add an Authorization header with value Bearer <your-key>

Open the model picker in VS Code Chat (Cmd+Shift+M) and your cloud models appear under the Ollama section.

Adding your harness

Each launcher is ~10 lines. Copy one, then:

1. Source the lib and require the key:

   source "$(dirname "$(readlink -f "$0")")/lib.sh"
   _require_key

2. Resolve the model — parses --model, else chooser/OLLAMA_MODEL:

   _resolve_model <harness-name> "$@"   # sets $MODEL and array $REST (leftover args)

   <harness-name> is just the key for the last-used file.
3. Wire the endpoint to https://ollama.com using $OLLAMA_API_KEY, then exec the tool with $MODEL and the passthrough args:

   exec yourtool --model "$MODEL" "${REST[@]+"${REST[@]}"}"