Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

239 — Agent Frameworks & Infrastructure Evaluation for ZAO OS

Status: research-complete Date: April 1, 2026 (v2 — refreshed star counts + revised verdicts) Topic: agents Type: infrastructure-research Last-validated: 2026-05-21 Original-query: Agent frameworks infrastructure evaluation for ZAO OS - OpenClaw, gstack, n8n, Firecrawl, ElizaOS, Paperclip, security/search/automation tools Tier: reference Goal: Evaluate 16 agent frameworks, security tools, search/data tools, and workflow automation for ZAO OS adoption Context: ZAO OS runs OpenClaw (ZOE agent) on a Hostinger VPS, uses gstack skills locally, and is operated by a 1-person team.


Live GitHub Data (Fetched April 1, 2026)

# Tool Stars License Last Updated
1 OpenClaw 373,689 MIT May 21 2026
2 n8n 188,989 Fair-code May 21 2026
3 Firecrawl 102,247 AGPL-3.0 Apr 1 2026
4 gstack 60,307 MIT Apr 1 2026
5 deer-flow (ByteDance) 55,776 MIT Apr 1 2026
6 Task Master AI 26,343 Proprietary Apr 1 2026
7 GPT Researcher 26,167 Apache-2.0 Apr 1 2026
8 promptfoo 18,981 MIT Apr 1 2026
9 elizaOS/eliza 18,007 MIT Apr 1 2026
10 claude-code-security-review 4,101 MIT Apr 1 2026
11 container-use (Dagger) 3,695 Apache-2.0 Apr 1 2026
12 CK (BeaconBay) 1,542 Apache-2.0 Mar 31 2026
13 Ghost OS 1,271 MIT Apr 1 2026
14 cmux 471 MIT Apr 1 2026
15 agent-governance-toolkit (MS) 352 MIT Apr 1 2026
16 figaro 117 MIT Mar 31 2026
17 Agent Alchemy 31 MIT Mar 31 2026

Recommendations Summary

# Tool Verdict Priority Why
1 OpenClaw USE P0 Already deployed. 344K stars, fastest-growing OSS ever. ZOE lives here.
2 gstack USE P0 Already installed. 60K stars. /qa, /ship, /review are daily drivers.
3 claude-code-security-review USE P1 Free GitHub Action. Every PR gets Claude security scan. 5 min setup.
4 n8n USE P1 Deploy on VPS. Automate ZOE triggers, cross-posting, fractal reminders.
5 container-use (Dagger) USE P2 Sandbox ZOE's code execution. Prevents rogue agent actions on VPS.
6 promptfoo WATCH P2 Red-team ZOE before exposing to community. Overkill now, critical at scale.
7 Firecrawl WATCH P2 Good for bulk web scraping. ZAO has Tavily + WebSearch already.
8 CK (BeaconBay) WATCH P3 Semantic code search. grep works at current scale.
9 Task Master AI SKIP -- Redundant with gstack + superpowers:writing-plans + autoresearch.
10 deer-flow (ByteDance) SKIP -- Python research agent. /zao-research + /autoresearch does this natively.
11 GPT Researcher SKIP -- Same as deer-flow. Existing research workflows are purpose-built.
12 cmux SKIP -- superpowers:dispatching-parallel-agents is the native approach.
13 figaro SKIP -- Fleet orchestration. ZAO needs 1 agent, not a fleet. 117 stars.
14 Agent Alchemy SKIP -- 31 stars. gstack is 2000x more adopted and already installed.
15 Ghost OS SKIP -- AI controlling Mac apps. Irrelevant to a web platform.
16 agent-governance-toolkit (MS) SKIP -- Enterprise multi-agent governance. ZAO has 1 agent.

Detailed Evaluations

TIER 1: Already Using -- Keep Investing

1. OpenClaw (344,026 stars, MIT)

What it does: Config-first agent framework. Write a SOUL.md, connect channels (Telegram, Discord, WhatsApp, 25+ more), plug in MCP servers, agent is live. No Python, no LangChain graphs.

ZAO status: ZOE agent deployed on Hostinger VPS (31.97.148.88). Running v2026.3.27. Needs SOUL.md, Telegram bot, and knowledge base loading (see doc 204).

Why it matters: 344K stars in ~60 days -- fastest-growing OSS project ever. Massive ecosystem: 162 agent templates, NemoClaw (Nvidia enterprise security), 25+ channel integrations. Config-first means Zaal iterates on ZOE's personality without writing framework code.

Next steps: Complete Phase 1-3 setup in doc 204. Priority: SOUL.md + Telegram + Supabase MCP.

2. gstack (60,307 stars, MIT)

What it does: Garry Tan's opinionated Claude Code skill suite. 23 tools acting as CEO, Designer, Eng Manager, QA, Release Manager, and Doc Engineer.

ZAO status: Fully installed at ~/.claude/skills/gstack/. Daily-use skills: /qa, /ship, /review, /design-review, /investigate, /plan-ceo-review.

Why it matters: Turns a 1-person team into a 6-person team output. 60K stars, 7.9K forks in 3 weeks. Already deeply integrated with ZAO's superpowers skill layer.

Next steps: Keep updated via /gstack-upgrade. No action needed.


TIER 2: Should Adopt Soon

3. claude-code-security-review by Anthropic (4,101 stars, MIT)

What it does: GitHub Action that runs Claude on every PR to find security vulnerabilities. Analyzes diffs for secrets, injection, auth bypass, etc.

Why USE: ZAO handles auth tokens, Supabase service keys, wallet operations, and user data. A free automated security reviewer on every PR is essential for a solo dev who cannot self-review blind spots. 5 minutes to add to .github/workflows/.

Cost: Free (GitHub Actions minutes).

Next steps: Add GitHub Action to ZAO OS repo. Combine with /autoresearch:security for deeper audits.

4. n8n (181,980 stars, Fair-code)

What it does: Visual workflow automation with 400+ integrations and native AI nodes. Self-hostable. Open-source Zapier with LLM support.

Why USE (revised from previous WATCH): At 182K stars and with Docker support, n8n fills a real gap in ZAO's stack -- the automation glue between systems:

  • Supabase row insert -> trigger ZOE response in Telegram
  • New governance proposal -> cross-post to Farcaster/X/Bluesky (replacing custom src/lib/publish/ code)
  • Music submission -> moderation pipeline -> notification
  • Fractal process reminders (Monday 6pm EST)

Cost: Free self-hosted. Fair-code license is fine for internal single-instance use.

Concern: Fair-code license prohibits redistribution as SaaS. Fine for internal ZAO use.

Next steps: Docker-compose on VPS alongside OpenClaw. Start with one workflow: fractal reminder automation.

5. container-use by Dagger (3,695 stars, Apache-2.0)

What it does: Containerized dev environments for coding agents. Each agent gets an isolated container with its own filesystem, tools, and network.

Why USE (revised from previous SKIP): ZOE runs on the same VPS as the OpenClaw gateway. Standard OpenClaw Docker provides process isolation, but container-use provides stronger sandboxing for ZOE's code execution -- separate filesystem, network controls, resource limits. Prevents a rogue agent action from corrupting the workspace or leaking data.

Cost: Free. Needs Docker (already installed on VPS).

Next steps: Evaluate after ZOE is fully configured (post doc 204 completion).


TIER 3: Watch List

6. promptfoo (18,981 stars, MIT)

What it does: Automated testing and red-teaming for AI models. Define test cases, run against prompts/agents, get pass/fail. Covers jailbreaks, prompt injection, data leakage, bias. Used by OpenAI and Anthropic themselves.

Why WATCH: ZOE is not yet community-facing. Once ZOE talks to members via Telegram/Discord, prompt injection becomes a real risk. promptfoo tests "can someone trick ZOE into revealing Supabase keys?" before going live.

When to USE: When ZOE's Telegram bot goes public to community members.

7. Firecrawl (102,247 stars, AGPL-3.0)

What it does: Turn any website into clean, LLM-ready markdown. Handles JS rendering, pagination, sitemaps.

Why WATCH: ZAO has Tavily + WebSearch + Jina Reader (doc 235). Firecrawl is better for bulk scraping (e.g., "scrape every track on this Bandcamp page"). AGPL license is viral -- modifications must be open-sourced.

When to USE: If ZOE needs structured data from music platforms at scale.

8. CK by BeaconBay (1,542 stars, Apache-2.0)

What it does: Semantic code search. Search codebase by meaning, not keywords. Local-first, uses embeddings.

Why WATCH: Claude's 1M context + grep + 164 research docs handles current scale. Revisit if codebase grows past ~100K LOC.


TIER 4: Skip

9. Task Master AI (26,343 stars)

PRD-to-task pipeline for AI coding agents. Skip: gstack /plan-ceo-review + /plan-eng-review + superpowers:writing-plans + /autoresearch already do this. Adding another task system creates confusion.

10. deer-flow by ByteDance (55,776 stars, MIT)

Python research agent with sub-agents and sandboxes. 55K stars is impressive. Skip: /zao-research + /autoresearch + 164 research docs is a purpose-built research system already tuned to ZAO's format.

11. GPT Researcher (26,167 stars, Apache-2.0)

Autonomous research agent producing compiled reports. Skip: Same reason as deer-flow.

12. cmux (471 stars, MIT)

tmux multiplexer for Claude Code. Skip: superpowers:dispatching-parallel-agents does this natively.

13. figaro (117 stars, MIT)

Fleet orchestration for Claude agents across containers, VMs, devices. Skip: ZAO needs 1 agent (ZOE), not a fleet. 117 stars = very early.

14. Agent Alchemy (31 stars, MIT)

Plugin collection for Claude Code. Skip: 31 stars, early. gstack (60K stars) is the same concept but vastly more mature.

15. Ghost OS (1,271 stars, MIT)

AI agents controlling native Mac apps (Finder, Safari, Calendar). Skip: ZAO is a web platform. Native Mac control is irrelevant.

16. agent-governance-toolkit by Microsoft (352 stars, MIT)

Zero-trust identity, execution sandboxing, policy enforcement. OWASP Agentic Top 10. Skip: Enterprise multi-agent governance at wrong scale for ZAO. /autoresearch:security handles audits.


Priority Action Plan

This Week (P0)

  • OpenClaw -- already deployed, continue doc 204 setup
  • gstack -- already installed, keep using

This Month (P1)

  • claude-code-security-review -- Add GitHub Action to ZAO OS repo (5 min)
  • n8n -- Docker-compose on VPS, build first workflow (fractal reminder)

Next Quarter (P2)

  • container-use -- Sandbox ZOE after full configuration
  • promptfoo -- Red-team ZOE before opening Telegram bot to community

Revisit Later (P3)

  • Firecrawl -- If ZOE needs bulk music platform scraping
  • CK -- If codebase grows past 100K LOC

Key Insight

ZAO's existing stack (OpenClaw + gstack + Claude Code skills + superpowers) covers ~80% of what these 16 tools offer. The real gaps are:

  1. Automated PR security -- claude-code-security-review fills this (P1)
  2. Workflow automation glue -- n8n fills this (P1)
  3. Agent sandboxing -- container-use fills this (P2)
  4. Agent prompt security testing -- promptfoo fills this when ZOE goes public (P2)

Everything else is either redundant with existing tools or solving problems ZAO does not yet have.


Changes from v1 (Previous Evaluation)

Tool Old Verdict New Verdict Why Changed
gstack USE (3.2K stars) USE (60,307 stars) Star count was stale; 19x growth confirms it
n8n WATCH (68K stars) USE (181,980 stars) 2.7x growth + AI nodes make it the right glue layer
container-use SKIP USE (P2) Stronger sandboxing than basic Docker for agent code execution
Firecrawl SKIP WATCH 102K stars, massive adoption; keep on radar for music scraping
deer-flow WATCH (4.1K) SKIP (55,776) Grew 14x but still Python-based; ZAO's research is JS-native
claude-code-security-review USE (1.2K) USE (4,101) 3.4x growth validates the approach

Sources