Installation¶

Supported Python Versions¶

The core package requires Python 3.10 or newer and advertises Python 3.10 through 3.14. Optional integrations can require a newer interpreter; this is not a promise that every extra runs on every core-supported Python version.

Install with uv¶

For a new application project, create an environment first:

uv init my-kedi-project
cd my-kedi-project
uv add kedi

Run Kedi inside the project environment:

uv run kedi --help
uv run kedi -c '= ready'

For a temporary CLI invocation without adding a project dependency, use uvx kedi --help. Use uv run kedi for the rest of this guide when using a uv project, or activate .venv so the shorter kedi commands resolve correctly. Commit uv.lock and use uv sync --locked in automation to preserve the tested dependency resolution.

Install with pip¶

Kedi can also be installed into an activated virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install kedi

Use uv for the repository's contributor workflow. The pip path is intended for consumers whose environments are managed by another tool.

On Windows PowerShell, activate with .\.venv\Scripts\Activate.ps1 instead. The source command above is for POSIX shells.

Configure Your First Model¶

No provider or API key is needed for parsing or deterministic Python expressions. The model-backed learning examples explicitly select openai:gpt-5.6-luna through the Pydantic adapter. Install its provider support in the same environment:

uv add 'pydantic-ai-slim[openai]'

For a pip-managed environment, use python -m pip install 'pydantic-ai-slim[openai]'. Keep Kedi installed so the resolver retains its compatible Pydantic AI bounds. Set OPENAI_API_KEY in your shell or a local, untracked .env file:

OPENAI_API_KEY=your-key-here

The kedi command loads dotenv at startup. The Python API does not imply the same CLI startup step: call dotenv.load_dotenv() yourself when embedding and depending on a .env file. A key authenticates the provider; it does not select a model. The source example supplies both > adapter: and > model:.

Do not commit .env, credentials, or captured secret-bearing output. A real model call can incur provider charges. Start with the no-model verification below.

Optional Backend Dependencies¶

The core distribution includes the Pydantic AI adapter. Other adapters, provider SDKs, and runtime surfaces are installed only when selected. These are alternative additions for different applications, not one installation sequence:

uv add "kedi[claude]"
uv add "kedi[codex-model]"
uv add "kedi[dspy]"
uv add "kedi[groq]"
uv add "kedi[langchain]"
uv add "kedi[langchain-aws]"
uv add "kedi[playground]"
uv add "kedi[notebook]"
uv add "kedi[typesafe]"

dspy installs the DSPy adapter and optimization instrumentation. langchain installs the LangChain adapter's common OpenAI, OpenRouter, and MCP integrations; add langchain-aws only for Bedrock. groq installs the provider SDK used by Groq-backed Pydantic AI models. claude and codex-model install their corresponding SDK bridges, while playground adds browser server dependencies. Codex, Claude, and ACP harnesses may also require an executable, authentication, or explicit command configuration.

Install only what the program uses. The Codex-model extra targets Python 3.11+ and Terminal-Bench tooling targets Python 3.12+. Notebook and Jev setup have their own guides: Notebook and Typesafe / Jev. Installing an extra does not authenticate a provider, start a harness, or change the model selection.

Verify the CLI¶

kedi --help
kedi -p -c "= ready"
kedi -c "= ready"

The second command checks syntax; the third executes and prints ready. Neither contacts a model. kedi-lsp is a stdio service launched by an editor, not an interactive CLI verification command. Continue with Your First Program.

Upgrade Kedi¶

With uv:

uv lock --upgrade-package kedi
uv sync

Review release notes before upgrading a production workflow. Kedi programs may also depend on provider SDK behavior, agent harness versions, and installed Kedi packages.