A Rust Layer for Agent Terminals That Shrink Token Bills

What it does

ANOLISA is a collection of Rust-based server-side tools that wrap around your existing shell, agent framework, and sandbox. Its best-known pieces are cosh-ng, an AI-native terminal that mixes Bash commands with natural language, and Token-less, a compression layer that sits between your agent and its LLM to prune redundant tool schemas and responses. The project also bundles observability, cross-session memory, checkpointing, and sandboxing under a single CLI installer.

The interesting bit

The Token-less compressor is the standout: it claims up to 62.9% fewer tokens in the full pipeline by blacklisting verbose fields like debug and trace, replacing dropped arrays with reversible <<tokenless:KEY>> markers, and doing it all in roughly 200 microseconds without touching the agent framework’s code. On Linux, AgentSight uses eBPF to trace token spend and tool calls kernel-side, which is more invasive instrumentation than you usually see in agent tooling.

Key highlights

• cosh-ng keeps familiar Bash/Zsh behavior but adds an agent that can use tools, invoke Skills, and ask for approval before risky operations.
• Token-less strips redundancy from tool responses and schemas; the README cites one workload where it saved 317K tokens (40.5%), though it notes results vary.
• AgentSight provides eBPF-based observability on Linux to trace user input, model calls, and token usage without modifying agent code.
• Agent Sec Core and SkillFS isolate risky operations and detect Skill drift—flagging changed, signed Skills as drifted before execution.
• Components are modular: you install the CLI entry point and enable only the capabilities you want.

Caveats

• The README notes that Token-less savings apply only to tool responses entering the context, not the whole session bill, and that its 40.5% figure comes from a single observed task with varying results.
• Several runtime components—such as ws-ckpt, Blaze, and ktuner—are listed in the component table but receive no detailed explanation in the README.
• The project describes itself as an “Agentic OS,” yet the sources emphasize keeping your existing shell, framework, and sandbox, suggesting it functions more as a modular middleware layer than a replacement operating system.

Verdict

Worth a look if you are running agent-heavy workflows from Linux or macOS and watching token costs climb. Skip it if you are on Windows or expect a fully integrated agent framework rather than a pick-and-mix middleware layer.