SAIL-Research-Lab/cheetahclaws

CheetahClaws, From Model Scaling to System Scaling: Scaling the Harness in Agentic AI

What it solves

CheetahClaws provides a hackable, Python-native alternative to production-grade AI coding assistants like Claude Code. It solves the problem of being locked into a single model provider or a compiled, opaque codebase, allowing developers and researchers to run agentic assistants using any LLM—including local, offline models—without needing a complex Node.js build chain.

How it works

The project implements a streaming API and an automatic tool-use loop (contained primarily in a single agent.py file). It connects to various model providers (Anthropic, OpenAI, Gemini, Ollama, etc.) and manages a suite of built-in tools for file manipulation, shell execution, and web searching. It features a multi-layer context compression system to handle long conversations and a dual-scope persistent memory system to track user and project-specific information.

Who it’s for

It is designed for developers who want a local or non-Anthropic coding assistant, AI researchers studying agentic loops, and teams needing a customizable baseline for building AI agents.

Highlights

  • Model Agnostic: Supports a wide array of providers including OpenRouter, Requesty, and local engines like Ollama and vLLM.
  • Extensible Tooling: Includes 28 built-in tools and supports the Model Context Protocol (MCP) for adding external tools via stdio, SSE, or HTTP.
  • Autonomous Execution: Features an "auto" mode that carries out requests end-to-end, with a granular permission system to prompt only for high-risk actions.
  • Developer Experience: Includes a terminal REPL with ghost-text predictions for the next prompt, a Web UI, and a desktop Electron app.
  • Advanced Agentic Features: Supports spawning typed sub-agents, task dependency graphs, and a "Brainstorm → Worker" workflow for persona-based planning and implementation.

Related

  • Project
    johnson7788/MultiUserClaw

    MultiUserClaw is a Docker‑based SaaS framework that lets you run isolated AI assistants (Hermes agents) for many users. It includes a React front‑end, a FastAPI gateway for auth, container management and LLM proxying, per‑user Hermes containers, PostgreSQL storage, and features like knowledge bases, skill store, cron jobs, multi‑channel bots, and multi‑model support.

  • Project
    clawdotnet/openclaw.net

    OpenClaw.NET is a .NET‑native, self‑hosted AI‑agent runtime with a desktop Avalonia UI (AgentQi Companion). It runs a gateway that talks to any OpenAI‑compatible model, offers streaming, tool execution, session memory, and many integrations (REST, WebSocket, MCP, 80+ tools, channel adapters). The gateway can be compiled with NativeAOT, making it suitable for edge or on‑prem deployments. Users can start via a ready‑made desktop bundle or from source, configure the assistant through chat commands, and extend functionality with .NET plugins or reusable MetaSkill DAG workflows.

  • Project
    zhayujie/CowAgent

    CowAgent is an open‑source, self‑hosted AI assistant that plans tasks, runs tools, keeps a personal knowledge base, and can operate as a team of cooperating agents. It supports many LLM providers, multimodal input, dozens of chat channels (Telegram, Slack, WeChat, etc.), a plug‑in skill system, and a three‑tier memory with nightly “Deep Dream” distillation. Install via a one‑line script or Docker, manage everything through a web console, and extend the system through the public Skill Hub or custom tool integrations.

  • Project
    bytedance/deer-flow

    DeerFlow is an open‑source framework that lets you build autonomous AI agents capable of orchestrating sub‑agents, using tools (web search, sandboxed code execution, RAG), and managing memory. It supports many LLM providers, offers a web UI, a terminal UI, and integrates tracing services, all deployable via Docker.

  • Project
    Trampoline-AI/predict-rlm

    predict-rlm is a Python library that gives large language models a sandboxed, stateful runtime. Models write and execute Python code, keep mutable state, call focused sub‑models, and use tools/skills (e.g., PDF handling, spreadsheets). The framework records a full trace for observability and can feed those traces to optimisation tools like GEPA. It’s aimed at complex, multi‑step workflows such as document analysis, audit, or spreadsheet automation, and supports optional Docker sandboxing and rich callbacks.