FollowAgents AI AGENTS DIRECTORY
Home / All Agents / ANOLISA Agent OS Layer
Automation & Ops terminal-interfacetoken-compressioncontext-managementworkspace-checkpointingruntime-isolationskills-mountingobservability

ANOLISA Agent OS Layer

A terminal, context-efficiency, and execution-layer toolkit for existing AI agent stacks.

View on GitHub ↗ Quick start ↓
Source repo
agentic-os-org/ANOLISA
Stars
★ 660
Last updated
today
License
Apache-2.0
Primary language
Rust
FA score
65/100 · Some gaps

At a glance

How it runs
CLI
Works with
Universal · cross-platform
Cost
Free software; you pay for model usage
Setup effort
Low · running in minutes
You'll need
Linux or macOScurlShell / CLINetwork accessLocal filesystem
Typical use
A platform team already running an agent framework and sandbox can add cosh-ng when its agents need a structured terminal interface for Shell and system work.
Not a fit if
  • Teams running Windows, since only Linux and macOS are supported
  • Production teams needing documented isolation and recovery semantics today
Source review
65/100 · Some gaps

What does this agent do, and when should you use it?

ANOLISA is a server-side operating layer for AI agent workloads, with ANOLISA CLI as its common installation entry point. Its cosh-ng component gives agents a structured, predictable interface for Shell and system operations. Token-less sits between an agent and its model to compress tool schemas and tool responses, while retaining references for removed array items. Agent Memory, SkillFS, and AgentSight cover cross-session context reuse, on-demand Skill mounting, and Token-spend recording. The execution-environment direction includes Agent Sec Core for isolating risky operations and ws-ckpt for workspace recovery points; the README identifies Linux and macOS as supported platforms.

After installation, operators can run anolisa install cosh-ng or anolisa install tokenless, then start the agent-native shell with cosh-ng. Token-less uses ResponseCompressor and SchemaCompressor on tool responses and schemas entering context; the documented example drops blacklisted debug and trace fields, null metadata, and empty tags / extra values. Removed array items retain a retrievable <<tokenless:KEY>> marker, making that compression reversible. Agent Memory reuses context across sessions, SkillFS exposes Skills as views and mounts them on demand, and AgentSight records where Tokens are spent. Agent Sec Core is described as isolating risky operations, while ws-ckpt preserves recovery points for workspace changes.

  1. A platform team already running an agent framework and sandbox can add cosh-ng when its agents need a structured terminal interface for Shell and system work.
  2. An agent application whose tool outputs contain debug data, nulls, empty fields, or redundant arrays can use Token-less before those responses enter model context.
  3. A team running recurring agent sessions can evaluate Agent Memory when it wants to reuse context across sessions.
  4. A developer organization with many Skills can use SkillFS when only the relevant Skills should be mounted into context for a given task.
  5. An operations team investigating context cost can use AgentSight to record where Tokens are actually spent.
  6. A workspace-oriented agent environment can assess ws-ckpt for recovery points and Agent Sec Core for risky-operation isolation.

How do you install or deploy this agent?

The documented installation command is:

curl -fsSL https://get.agentic-os.sh | bash

Install the desired components:

anolisa install cosh-ng
anolisa install tokenless

The README identifies Linux and macOS as platforms. This installation path requires network access and curl; no API key, account, or other credential requirement is documented.

How do you use this agent?

After installing cosh-ng, run:

cosh-ng

This enters the agent-native shell. The README also says an existing agent can be pointed at cosh-ng. Once Token-less is installed, it applies to tool calls; the supplied documentation does not provide an additional configuration command, so validate compression behavior and retrieval-marker handling in the target workflow.

What are this agent's strengths and limitations?

Pros
  • It is designed to retain the Shell, agent framework, and sandbox already in use instead of replacing the execution stack.
  • Token-less covers both tool schemas and tool responses, while preserving retrievable <<tokenless:KEY>> references for removed array items.
  • Its context-efficiency components span session reuse, on-demand Skill mounting, and Token-spend recording rather than only response compression.
  • The execution-environment scope includes both risky-operation isolation and workspace recovery points.
Limitations
  • The README explicitly identifies only Linux and macOS; it provides no evidence of Windows support.
  • The published Token-reduction figures apply to tool responses and schemas entering context, not proportionally to the entire session bill.
  • Initial setup executes a remotely fetched script through curl, which requires network access and may need to pass organizational software-installation review.
  • Although it is described as working with existing frameworks and sandboxes, no named framework, model provider, or platform integration is documented, so compatibility needs validation.
  • Agent Sec Core and ws-ckpt are presented as execution-environment capabilities being built out, without documented isolation semantics, recovery procedures, or configuration details.

How does this agent compare with similar options?

ANOLISA is positioned as an added layer around an existing Shell, agent framework, and sandbox, rather than as a replacement for those systems.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
ANOLISA Agent OS Layer This agent 65 · Some gaps CLIFree + model costs ★ 660 today Rust —
Arize Phoenix 70 · Some gaps CLIFreemium ★ 12k 1d ago Python OpenAI API · Claude API
cost-xray 65 · Some gaps CLIFree ★ 5.8k 3d ago Python Codex · Claude Code · OpenAI API · Claude API
Adrian 59 · Major gaps Library / SDKFreemium ★ 589 18d ago Python Claude Code · OpenAI API · Claude API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Some gaps
65/ 100 5-point scale 3.3 / 5
This review covers the source as of Nov 08, 2026. The project has shipped 2 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 18/29
Reliability 6/14
Adaptability 15/18
Convention 12/18
Effectiveness 9/13
Verifiability 5/8
Why each dimension lost points
Trust18 / 29 · 3.1/5

The PII Checker limits scans to user-designated content and prefers files or stdin to avoid placing sensitive text in command arguments. The policy-authoring skill requires clarification, self-checking, and a natural-language readback, and explicitly says it does not approve or deploy policies, supporting strong least-privilege, confirmation, and sensitive-data handling. The README describes sec-core and underlying isolation and workspace-checkpoint rollback. The supplied material does not show dependency auditing or a complete data-flow account, and publisher identity is unverified, so dependency security and attribution receive only partial support.

Reliability6 / 14 · 2.1/5

CLI examples describe parameters, structured results, and error interpretation; the memory service also logs startup, quota failure, and shutdown errors. However, the Prompt Scanner's exit-code and schema descriptions are not fully consistent (for example, its sample has ok=false with verdict=warn, while the prose associates ok with no threat), and standard mode may silently degrade to L1-only scanning with a warning. This limits self-consistency and dependency-availability scores. Failure handling is described concretely, though evidence is not equally detailed for every component.

Adaptability15 / 18 · 4.2/5

The README addresses terminal users, Agent developers, platforms, and infrastructure through several integration paths. Skills distinguish code, PII, prompt scanning, and policy authoring, and specify input scope, platform conditions, and policy limits. Trigger descriptions are generally clear, but the Prompt Scanner trigger is broad, including pre-execution checks and batch audits, which may cause over-triggering. Platform and installation requirements are substantial, but vary across components and external services, so environment fit is not fully unified.

Convention12 / 18 · 3.3/5

The README provides a clear component architecture, repository map, and user/developer entry points; it also links installation guidance, examples, licensing, and some limitations. Examples cover multiple scenarios, but FAQs and limitations are dispersed, and installation depends on external guides with component-specific requirements. The repository's rule to preserve package names and release tags supports naming stability, though the supplied sample has a potentially stale name (the security policy still lists copilot-shell among supported packages). A changelog is linked but version history is not present in the sample; contribution channels exist, while named maintenance responsibility and update commitments are unclear.

Effectiveness9 / 13 · 3.5/5

The skills define structured JSON fields, verdicts, and user-facing interpretation; the policy-authoring skill also requires translating generated rules back into plain language. These features make outputs useful. The README describes concrete intended benefits in compression, memory, security, and observability, but the supplied evidence mainly offers product claims and demonstration links, without independently verifiable quantitative gains or a cost tradeoff. Marginal value and cost-benefit therefore remain only partly supported.

Verifiability5 / 8 · 3.1/5

The README links component capabilities to installation paths and reproduction guides; skills specify inputs, output fields, verdict interpretation, and limitations, while the policy skill identifies bundled schema, syntax, and self-check references. Some claims are tied to concrete files or examples, but many supporting materials are outside the supplied packet, and product capabilities lack independent result evidence here. The text does distinguish scan outcomes from security guarantees and policy self-checks from deployment, but the boundary between fact and inference is weaker for performance and cost claims.

Risks and how to mitigate them
  • The Prompt Scanner's schema prose and examples are not fully consistent. Standard mode may run only the rule layer when ML dependencies are unavailable, so users should inspect the actual verdict, layer results, and warnings.
  • The supplied material does not quantify Tokenless cost savings; documentation for policies and scanners does not substitute for runtime validation.
  • Components have differing platform, service, and installation requirements. The security policy's supported-version table may contain a legacy component name; check the relevant component documentation before adoption.
Evidence confidence: Low Reviewed Nov 08, 2026 Reviewed revision 23e3dfc0bc6e New commits since this review; the score may not cover them
Review evidence README.mdAGENTS.mdcode-scanner/SKILL.mdpii-checker/SKILL.mdpolicy-authoring/SKILL.mdprompt-scanner/SKILL.mdsrc/main.rssrc/main.rssrc/main.rsLICENSESECURITY.mdworkflows/_rpm-build.yamlworkflows/agent-memory-prebuilt-ci.yamltests/__init__.pytests/conftest.pytests/helpers.py
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/anolisa.svg)](https://followagents.com/en/agents/anolisa)
<a href="https://followagents.com/en/agents/anolisa"><img src="https://followagents.com/badge/anolisa.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. memory/v0.2.14 agent-memory v0.2.14
    Release notes excerpt
    Agent Memory v0.2.14 Changed - Breaking for source installations: make install now defaults to AWPRJ=anolisa, placing the adapter in the shared ANOLISA namespace while preserving unrelated files in that adapter tree. To keep managing an existing independent installation, pass AWPRJ= explicitly; to migrate, first run make uninstall AWPRJ= with the original profile and path overrides. This source-install change leaves runtime configuration paths and RPM/raw payload layouts unchanged (#9462). …
    View on GitHub ↗
  2. anolisa/v0.4.0 anolisa v0.4.0
    Release notes excerpt
    ANOLISA CLI v0.4.0 Added - Add the internal anolisa telemetry prepare command for image builds. It prepares telemetry files and enables the resident systemd unit for the next boot without starting collection or recording the build host's instance snapshot. Applying it requires root and system scope; an existing opt-out marker or service setup failure stops preparation. Use --no-service to skip unit setup, then run anolisa telemetry init at boot before anolisa telemetry upload --loop. Existing RPM and image-build flows are not …
    View on GitHub ↗
  3. cosh-ng/v0.29.0 cosh-ng v0.29.0
    Release notes excerpt
    Cosh NG v0.29.0 Added - Alibaba Cloud ECS instances can now answer questions without /auth: eligible requests use the OS Console model access automatically, defaulting to qwen3.7-plus, while an already configured provider stays selected; /auth gains an identity entry to switch back to the saved identity or to another provider, n disables automatic access for the current or future sessions, and Esc cancels the activation wait without replaying the question (#9449) …
    View on GitHub ↗

FAQ

Do Token-less savings equal the total model-session cost reduction?
No. The README states that the reported savings apply to tool responses and schemas entering context, not to the full session bill.
Does Token-less require changes to agent-framework code?
The README says compression runs between the agent and model, so no agent-framework code changes are required.
Is compressed information permanently lost?
The README says removed array items retain a retrievable <<tokenless:KEY>> marker, intended to make compression reversible.
What credentials are required for installation?
No API key, account, or other credential requirement is documented. The shown installation command requires network access and curl.
Which operating systems are explicitly supported?
The README platform badge identifies Linux and macOS.
Was this page helpful?
View on GitHub ↗ Install ↓