ClydeCLI

Install, configure, run

Three steps. I'll wait. Might light one while you do it.

Install

One line (installs uv if needed, then clyde, then runs clyde setup):

curl -fsSL https://raw.githubusercontent.com/dea6cat/ClydeCLI/main/install.sh | sh
# or pin a release tag or an exact commit:
curl -fsSL https://raw.githubusercontent.com/dea6cat/ClydeCLI/main/install.sh | CLYDE_REF=v0.2.5 sh

Or with uv — the engine under the curl line, for when you already have it:

uv tool install --python 3.14 git+https://github.com/dea6cat/ClydeCLI
# pin a release: ...ClydeCLI@v0.5.11
# or try Clyde without installing:
uvx --from git+https://github.com/dea6cat/ClydeCLI clyde

Then run clyde setup once, as the curl line does. Or with pipx (needs Python 3.12 to 3.14, so name one if your default is older or newer; add --backend pip if pipx complains about an older uv). Python 3.11 is not supported: the bundled skill scanner, SkillSpector, needs 3.12.

pipx install --python python3.13 git+https://github.com/dea6cat/ClydeCLI
# pin a release: ...ClydeCLI@v0.2.5

Both give you clyde and bonnie, and clyde update and clyde uninstall use the tool that installed Clyde.

It includes SkillSpector, the bundled skill scanner, and the install is about 180 MB. Laya, the local decision model, is an optional extra (about 640 MB of packages, mostly PyTorch, plus its 800 MB model); an update keeps it if you already have it. clyde setup connects a provider and picks a default model, asks before adding Laya and downloading its model, lists the hooks, MCP servers and plugins you already set up for other agents (Claude Code, Cursor, Gemini CLI, Codex, Copilot CLI) and imports them only if you say yes, and offers to put clyde on your PATH. Pass --yes for no prompts (hooks, MCP servers and plugins are never imported that way). Run it again any time.

Or from source:

git clone https://github.com/dea6cat/ClydeCLI.git
cd ClydeCLI

# Create venv (uv recommended)
uv venv --python 3.14
source .venv/bin/activate

# Install
uv pip install -r requirements.txt

Configure

Where to get an API key for each provider, and a one-minute tutorial for adding it: dea6cat.github.io/ClydeCLI/keys.html.

Every feature with recordings, including tuning a local model (clyde tune: what it tries, the checks a setting must pass, real results): dea6cat.github.io/ClydeCLI/features.html

Option 1: Just export a key (no login needed)

If a provider key is already in your environment (see the table above), run clyde. With no default model saved, it picks one from whatever is connected (local Ollama first), tells you which, and /model switches it.

Option 2: Interactive login

python -m src.cli login

This flow will:

  1. ask you to choose a provider
  2. ask for that provider's API key (shown as * while you type or paste)
  3. fetch the provider's live model list and ask for a default model
  4. save the key to ~/.clyde/keys.json and the model to ~/.clyde/config.json

For Ollama there's no key: it checks the server is up, and if you have no models yet it suggests tool-capable ones that fit your RAM.

Inside the REPL, /login runs the same flow and switches to the model you pick (/login nvidia skips the provider question, handy for replacing a key).

A provider that isn't built in: pick custom and give it a name, the base URL of any OpenAI-compatible API (the part before /chat/completions, e.g. https://api.together.xyz/v1) and its key (leave it empty for a keyless local server such as vLLM). It's saved under "providers" in ~/.clyde/settings.json, its key in keys.json (or export CLYDE_<NAME>_API_KEY), and from then on it works like a built-in one: /models, /eval, <name>:<model> and cardShuffle. clyde logout <name> removes both its key and its entry in settings.json.

Option 3: Pick the model per session

clyde --model openai:gpt-5.4 overrides the default for one session. A key exported in your shell always wins over a saved one.

The config file only holds the default model and session settings:

{
  "model": "anthropic:claude-sonnet-4-6",
  "session": {"auto_save": true, "max_history": 100}
}

Upgrading from an older config with per-provider api_key entries? The first run moves the keys to keys.json and keeps your default provider's model.

Run

python -m src.cli          # Start REPL
python -m src.cli --help   # Show help

That's all it takes: clone, configure, run.

Edit this page on GitHub