Agent infrastructure is evolving from "API calls in a loop" to "operating system" — Alibaba's OpenSandbox and Google Workspace CLI are surging simultaneously, signaling that big tech is betting agents need their own sandbox and toolchain rather than piggybacking on human dev environments.
General-purpose sandbox platform for AI applications. Multi-language SDKs, unified sandbox APIs, Docker/Kubernetes runtimes for Coding Agents, GUI Agents, Agent Evaluation, AI Code Execution, and RL Training.
| Stars | 6,586[exact] |
| 30d Growth | +4,059[exact] |
| 7d Growth | +3,556[exact] |
| Language | Python |
| Created | 2025-12-17 |
Why it matters — For agents to reliably execute code, they need isolated sandbox environments. E2B validated this path; now Alibaba enters with an enterprise-grade approach (Docker/K8s native). Burst ratio is only 2.33 — this isn't a hype spike, it's demand-driven sustained growth.
Paradigm signal — The L2 Runtime layer is evolving from "LLM + tool-calling" to "LLM + operating system." Sandboxes aren't just about security isolation — they're standardizing agent execution environments. This means an agent's capability boundary will be defined by its sandbox, not its prompt.
Google Workspace CLI — one command-line tool for Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin, and more. Dynamically built from Google Discovery Service. Includes AI agent skills.
| Stars | 14,405[estimated] |
| 30d Growth | +5,000+[estimated] |
| 7d Growth | +5,000+[estimated] |
| Language | Rust |
| Created | 2026-03-02 |
Why it matters — Google official, 14K stars in 4 days. The key isn't the CLI itself — it's the built-in AI Agent Skills. Google is signaling that the future Workspace interaction model is agent-driven CLI, not web clicking. The Rust implementation says this isn't an experiment.
Paradigm signal — Big tech is turning "AI Agent" from concept into product. Google choosing CLI over chatbot suggests they believe an agent's best interface is structured commands, not free-form conversation. This has directional implications for agent product design.
Agent Execution Environment Standardization
OpenSandbox (Alibaba), E2B/infra (open source), and the OpenClaw ecosystem (guardian, backup, cloud) are all active simultaneously. Agents are no longer just "API-calling loops" — they need OS-level infrastructure: sandboxes, state management, security boundaries, recoverability. This is the most important structural shift in the L2 Runtime layer.
MCP Ecosystem Rapid Expansion
IBM mcp-context-forge (unified gateway), ForLoopCodes/contextplus (semantic code graph MCP), AI-QL/tuui (desktop MCP client) — MCP has moved from protocol specification into toolchain competition. IBM's enterprise gateway signals MCP is transitioning from developer toy to production infrastructure.
OpenClaw Ecosystem Explosion
10+ new OpenClaw peripheral projects this week: skills collections, security guide (by SlowMist), Chinese translations, Android ports, backup tools, visual panels. Heavy L5 noise, but SlowMist's security guide and Overture's visual execution planner are worth noting — the former shows security teams are seriously auditing agent safety, the latter addresses agent explainability.
slowmist/openclaw-security-practice-guide, SixHq/Overture, cft0808/edict
| Repo | Layer | Stars | Summary |
|---|---|---|---|
| RightNow-AI/openfang | L2 | 11,648 | "Agent OS" concept, spike+decay growth — bold vision, watch for follow-through |
| katanemo/plano | L2 | 5,876 | AI-native proxy layer (routing + safety + observability), steady growth |
| BlockRunAI/ClawRouter | L2 | 4,969 | LLM router: 41+ models, <1ms routing, crypto payments on Base & Solana |
| IBM/mcp-context-forge | L2 | 3,371 | IBM's unified MCP/A2A gateway — enterprise adoption signal |
| cft0808/edict | L2 | 4,612 | 9-agent orchestration system with real-time dashboard and audit trails |
| AlexsJones/llmfit | L3 | 12,192 | One command to find what LLM runs on your hardware — model fragmentation is real |
| Panniantong/Agent-Reach | L3 | 6,546 | Give agents eyes: read Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu |
| Helicone/helicone | L3 | 5,198 | LLM observability platform, YC W23, one-line integration |
| maximhq/bifrost | L3 | 2,754 | Claims 50x faster than LiteLLM — enterprise AI gateway in Go |
| rowboatlabs/rowboat | L4 | 9,052 | AI coworker with memory, TypeScript |
| superset-sh/superset | L4 | 5,332 | IDE for the AI Agents Era — run multiple Claude Code/Codex simultaneously |
| open-pencil/open-pencil | L4 | 1,811 | AI-native design editor, open-source Figma alternative |
L1 Model/Inference (12) — Model-level and inference engines — slow-moving but foundational
maderix/ANE (5.9K, Apple Neural Engine reverse-engineering), gpustack/gpustack (4.6K), GLM-5 (1.7K), picolm (1.3K), MOSS-TTS (818), Helios (687), ssd (660), timber (599), doc-to-lora (472), AutoGrad-Engine (328), microgpt-c (230), picogpt (112)
L5 Wrapper/Demo (16) — Tutorials, skill collections, translations, awesome-lists — noisy but reflects ecosystem heat
claude-code-guide (3.5K), pm-skills (3.5K), OpenClawChineseTranslation (2.4K), AgentGuide (2.1K), OBLITERATUS (1.9K), openclaw-master-skills (1.4K), slowmist security guide (1.3K), openclaw-guardian (998), best-prompts (859), awesome-openclaw-usecases-zh (845), SwiftUI-Agent-Skill (695), excalidraw-diagram-skill (669), openclaw-backup (644), evals-skills (621), ui-design-brain (490), openclaw-docs (497)
Non-AI (4) — Caught by high-star strategy but not AI projects
MinecraftConsoles (3.5K), vphone-aio (989), AimAssist (68), higgsfield-ai-scam (48)