Installer llm avec Homebrew, apt, Nix

Liens du projet

• Page d'accueil
• Dépôt
• Docs amont
• Page du gestionnaire de paquets

Commandes et alias

• llm

Historique du projet

The changelog records version 0.1 on April 1, 2023 as the initial prototype release. Version 0.5 on July 12, 2023 added a plugin mechanism for additional language models, a key change that moved LLM beyond a single-provider OpenAI wrapper.

During 2023 and 2024, LLM added chat, embeddings, templates, SQLite logging and search, model aliases, attachments, async models, and broader provider support. In May 2025, version 0.26 added tool support, letting models execute Python functions through the CLI and Python API.

Historique d'adoption

LLM became part of the Datasette-adjacent command-line culture around small composable tools, SQLite-backed logs, and plugin systems. Homebrew, Debian, and Nix packaging made it easy to install as a normal developer utility rather than only as a Python package.

Its adoption expanded with the growth of provider-specific plugins and local-model integrations, giving users one command-line interface across OpenAI, Anthropic, Gemini, Ollama-backed models, and many community plugins.

Modes d'utilisation

Common uses include running one-off prompts from the shell, piping files into a model, starting interactive chats, storing API keys, saving prompt templates, logging responses to SQLite, generating embeddings, and calling models from Python code.

The documented configuration paths and keys.json locations matter to package users because Homebrew installs the executable, while user-specific model keys, templates, logs, and extra model definitions live outside the package prefix.

Pourquoi les passionnés de paquets s'y intéressent

LLM is package-nerd significant because it turns rapidly changing AI APIs into a stable Unix-style command with plugins. The packaging story is unusually important: users want a single binary-ish command in PATH, but the real extension surface is Python plugins and user configuration.

It is also a useful example of modern CLI state management: credentials, templates, provider models, logs, and tool definitions are intentionally external to the package, so upgrades can move the application forward without overwriting user data.

Chronologie

• 2023-04-01: LLM 0.1 initial prototype release.
• 2023-07-12: LLM 0.5 added the plugin mechanism for additional language models.
• 2024-10-29: LLM 0.17 added attachment support for multimodal models.
• 2025-05-27: LLM 0.26 added tool support.
• 2026-06-09: LLM 0.32a3 appeared in the changelog with tool-call and human-in-the-loop improvements.

Related projects

• Datasette is related through the broader Simon Willison/Datasette tooling ecosystem and the io.datasette.llm application data namespace.
• LLM plugins such as provider adapters and local-model integrations are central to the project's architecture.
• SQLite is related through LLM's prompt and response logging features.

Sources

• https://formulae.brew.sh/formula/llm
• https://github.com/simonw/llm
• https://github.com/simonw/llm/releases/tag/0.1
• https://llm.datasette.io/
• https://llm.datasette.io/en/stable/changelog.html
• https://llm.datasette.io/en/stable/setup.html