AnythingLLM

AnythingLLM is a desktop AI workspace for chatting with documents, choosing local or remote models and running agent workflows. It suits individuals who want a configurable document assistant on their own computer, with the option to connect external inference, embedding, search and vector-database services.

Documents, chat and agents

Import supported documents into workspaces and use retrieval to supply relevant context to model responses. Select an LLM, embedding model and vector database according to your workflow. Supported integrations include local model engines and cloud providers; the application's default local configuration should not be confused with every possible integration staying on-device.

Agents can use configured tools and skills for activities such as web browsing, file operations and document generation. Agent flows combine steps into repeatable workflows. Desktop-specific features include meeting and desktop assistants, while other deployment editions have their own capabilities. The upstream project also offers self-hosted and hosted deployments; their multi-user, network-access and administration behavior is not automatically the same as the Mac desktop application.

Install, hardware and dependencies

Download the correct official DMG for Apple Silicon or Intel, open it and drag AnythingLLM into Applications, or use brew install --cask anythingllm. The Mac installation guide explicitly distinguishes the two chip families and says Apple M-series devices run local inference faster. The checked Homebrew cask selects a Silicon DMG on ARM Macs and a separate DMG on Intel Macs. Neither should be described as a verified Universal installer. The checked installation and cask sources do not establish one specific minimum macOS version, so a numeric minimum is not invented here.

The publisher recommends 16 GB RAM and an eight-core CPU for a basic local-model experience; these are recommendations, not a promise that every model fits. Disk needs depend strongly on downloaded model weights and imported documents. Remote LLMs reduce local inference and model-storage requirements but require provider access. The cask declares no separate formula dependency or named conflict; local models, external servers and selected providers are additional runtime resources. Docker is not required merely to install the desktop DMG.

Grant microphone, screen or file access only for the optional assistant and import workflows you choose. No blanket Accessibility, Full Disk Access or administrator requirement is asserted for ordinary document chat. Imported agent skills and external MCP tools can have wider effects than chat alone; review their permissions and source before use.

Costs, network and accounts

The base desktop application and public project are available free, with MIT-licensed source. Desktop Pro and hosted services are separate paid offerings, while cloud-model providers can charge for usage and require API credentials. Model weights and external services have their own licenses and terms; the project's MIT license does not grant unrestricted rights to every model or connected dataset. Local workflows do not inherently require a hosted LLM account, but Pro, community and selected cloud integrations may require their own account or key.

Network access is used for downloads, updates, connected providers, web tools and optional services. Local models need adequate hardware and downloaded weights. Do not assume all agent actions work offline simply because a local model is selected.

Privacy and data risks

The desktop privacy policy says messages, chat histories and documents are saved locally by default, and documents optional anonymous usage telemetry that can be disabled in settings. That statement is about the default desktop workflow; choosing cloud inference, remote embeddings, online search or external tools can send the necessary content to those services. Review each provider's terms before processing confidential material.

AI responses and citations can be incomplete or wrong. Verify important conclusions against original documents. Keep backups of workspace data and source files, protect API keys and review agent file changes before execution. A local application is not itself a guarantee of confidentiality for every connected workflow.

HTTP HEAD on the official Silicon DMG reports 566,665,700 bytes for the checked download. Intel installers and installed model storage have different sizes.

Sources: Official product, Official project, Mac installation, Hardware recommendations, Desktop privacy, Official releases.