Cellar is an open-source project centered on the AI agent runtime. It is officially positioned as the "Computer Experience Layer" and described as an MCP server on macOS that supports native computer use, with structured perception, reliable operation, and continuous awareness capabilities.
In terms of purpose, CEL (the open portion) is an open context and trust data plane for AI agents, intended to provide agent builders with reusable contracts covering the four types of capabilities required by a serious agent runtime: fusing multiple sources into a canonical context snapshot; persisting cross-turn memory locally or in one's own backend; assembling a governable model brief with receipts; and inspecting what the agent sees, remembers, sends to the model, and later claims. In terms of main capabilities, according to the materials, its contracts are: multiple sources can emit or be captured as a unified ContextSnapshot; store and retrieve memory through the MemoryProvider trait; build a governed prompt bundle and inspect its receipts; understand action receipts and brief receipts; and connect context, memory, and brief assembly end to end. The project also mentions combining the accessibility tree, Chrome DevTools Protocol, and vision to form structured context, and supporting offline operation through Ollama.
In terms of usage, the current release line is 0.2.0 on crates.io; to upgrade from 0.1.x, see docs/migration-0.2.md. The materials explain that you can clone an individual crate repository separately, or run it in this workspace (cellar-oss). The top-level examples are organized by "task to be completed": examples/merge-context emits or captures multiple sources into a ContextSnapshot; examples/memory-provider stores and retrieves memory through the MemoryProvider trait; examples/build-brief builds a governed prompt bundle and inspects its receipts; examples/receipt-inspection understands action and brief receipts; examples/context-to-brief connects context, memory, and brief assembly end to end. For the guided path, you can read docs/quickstart.md; for the complete crate matrix, see docs/crates.md. Note: the specific installation commands and code content are not provided in the materials.
In terms of applicable scenarios, it is suitable for developers and builders who want to implement unified context, cross-turn memory, governed model briefs, and auditable receipts for agent runtimes.
In terms of limitations, this open-source project is intentionally not the complete Cellar runtime; the real-time cortex engine, policy enforcement, monitoring, compliance workflows, hosted workers, and GUI interface belong to the commercial Cellar/Dilipod operations layer built on these contracts. The open-source CEL crate and documentation use the Apache-2.0 license (unless a subdirectory states otherwise). Regarding platform limitations, the materials only explicitly mention native computer use on macOS; whether other platforms are supported is not yet provided.
