Anolisa – Agentic OS with runtime, security, observability and token compression Alibaba released ANOLISA, an open-source agentic operating system layer for AI Agent workloads, featuring a terminal interface (cosh-ng), token compression (Token-less), observability (AgentSight), and runtime security. In one observed coding task, Token-less saved 317K tokens (40.5%) based on AgentSight measurements, with tool responses reduced by 65.8% and tool schemas by 47.3%. A gentic N exus O perating L ayer & I nterface S ystem A rchitecture The operating system layer for Agent workloads. Let Agents drive the system straight from your terminal, and strip the tool responses that reach the model before they cost you — while keeping the Shell, Agent framework, and sandbox you already run. 中文版 /alibaba/anolisa/blob/main/README zh.md · Website https://agentic-os.sh/ · Quick Start https://agentic-os.sh/docs/quickstart/ · User Guide https://agentic-os.sh/docs/user-guide/ · Contributing https://github.com/alibaba/anolisa/blob/main/CONTRIBUTING.md ANOLISA is a server-side operating layer for AI Agent workloads. It addresses three practical constraints of Agent execution: terminal entry, Token cost, and execution environments. Keep the Shell, Agent framework, and sandbox you already use. ANOLISA CLI provides a single installation entry point, while each capability can be enabled independently. New to ANOLISA? Choose your first outcome in the Quick Start → https://agentic-os.sh/docs/quickstart/ | Agent entry | Context efficiency | Runtime & security | |---|---|---| Shell copilot | Tool-output compression | Checkpoint and rollback | System and DevOps expertise | Trace and Token visibility | Focused Skill views | Kernel tuning | Cross-session memory | Sandbox and verification | Sandbox lifecycle | 01 · AGENT INTERFACE cosh-ng is an AI-native Linux terminal: it keeps familiar Bash/Zsh behavior, then adds an Agent that can understand intent, use tools and Skills, and ask for approval before risky work. Shell commands and natural language share one terminal instead of forcing users into a separate chat application. 02 · CONTEXT EFFICIENCY Token-less https://agentic-os.sh/docs/user-guide/token-saving/tokenless/quickstart/ removes redundancy from tool schemas and responses before they reach the model. Agent Memory https://agentic-os.sh/docs/user-guide/token-saving/agent-memory/ reuses useful context across sessions. SkillFS https://agentic-os.sh/docs/user-guide/runtime/skillfs/ keeps the current Skill view focused and makes other Skills discoverable when needed. AgentSight https://agentic-os.sh/docs/user-guide/agent-observability/agentsight/ shows where Tokens are spent. On Linux, AgentSight uses eBPF to observe an Agent without changing its code. Follow user input through model and tool calls, with Token use and sub-agent branches in the same view. agentsight-demo-en-ebpf-cover-1080p.mp4 | Install Token-less and connect it to Claude Code: curl -fsSL https://get.agentic-os.sh | bash export PATH="$HOME/.local/bin:$PATH" anolisa install tokenless anolisa adapter enable tokenless claude-code Restart Claude Code, run one tool-heavy task, then inspect the result: tokenless stats summary tokenless stats list --limit 5 Open the full Token-less Quick Start → https://agentic-os.sh/docs/user-guide/token-saving/tokenless/quickstart/ · Read the user manual https://agentic-os.sh/docs/user-guide/token-saving/tokenless/user-manual/ tokenless-context-efficiency-hd.mp4 | In one observed coding task, Token-less saved 317K Tokens 40.5% , based on AgentSight measurements. Results vary by workload. debug and trace are dropped by the field blacklist, metadata as null, and tags / extra as empty values. Compression runs between the Agent and the model, so no Agent framework code changes. Dropped array items stay retrievable through a <