Five Open-Source AI Tools Dissected: From Toolchains to a Systems Ecosystem

> Originally published 2026-05-02. Translated and edited from zhichai.net.

These five projects seem unrelated at first glance—a Codex enhancement layer, a mobile AI showcase, a memory system, a self-evolving agent, and a screen-recording tool. But viewed together, they assemble the same picture: AI is moving from "conversational apps" to "engineering infrastructure."

Overview

| Project | Core Problem | Solution | |---------|-------------|----------| | OMX | Codex CLI lacks workflow and persistent state | Four-layer architecture: Skills + .omx/ state + tmux parallelism + AGENTS.md orchestration | | Edge Gallery | Cloud AI depends on network; limited privacy and latency | LiteRT on-device inference + Gemma 4 + FunctionGemma function calling | | MemPalace | Extractive memory is lossy and API costs are high | Verbatim storage + zero-LLM writes + progressive recall | | Hermes Agent | Agents lose memory and can't learn from experience | Built-in learning loop + self-generated skills + Honcho user modeling | | OpenScreen | Screen-recording post-production tools are expensive and closed | Electron + PixiJS + WebCodecs open-source pipeline |

1. OMX: Turning Codex from a "Chat Programmer" into an Engineering Team

OMX is a workflow enhancement layer for Codex CLI, not a replacement. Codex is smart but not engineering-ready—no requirements clarification, no persistent memory, no parallelism. OMX addresses these with a keyword-driven flow ($deep-interview → $ralplan → $ralph / $team) and persistent .omx/ state.

The most interesting engineering choice: tmux parallel workers + isolated git worktrees. Each worker is a full Codex process running in its own worktree, coordinated by a leader via CLI API—turning conversational AI into an engineering pipeline.

Maturity: ~27k stars, iterated to v0.15.0 in two months, maintained by a Korean team, MIT licensed. Best for developers using Codex CLI on medium-to-large projects.

2. Edge Gallery: Google's Showroom for On-Device AI

This is not a toy app—it's Google's on-device AI showcase, bundling the LiteRT inference engine, the Gemma 4 model family, and FunctionGemma function-calling into your phone.

On-device AI's value isn't just "no network"—it's the only solution for offline environments, privacy-sensitive scenarios, and real-time responses. LiteRT is a complete rebuild of TFLite, unifying CPU/GPU/NPU abstraction; NPU inference can reach 25× CPU speed at 1/5 the power.

Highlights:

• FunctionGemma 270m: a 270M-parameter function-calling specialist that runs offline on phones
• Mobile Actions: offline device control driven by FunctionGemma
• Thinking Mode: Gemma 4 supports both "thinking" and "non-thinking" modes with explicit reasoning traces

Maturity: 22.4k stars, official Google project, Apache-2.0, experimental Beta.

3. MemPalace: The Contrarian Bet on "Store Everything Verbatim"

Industry consensus is that LLMs should extract and summarize key facts (Mem0, Zep, LangMem all do this). But extraction is lossy—you don't know what will be asked in the future. MemPalace stores conversation transcripts untouched and relies on retrieval.

Its LongMemEval R@5 reaches 96.6% with pure semantic search and zero LLM calls, well ahead of leading extractive systems (Mem0 was ~49% at the time)—at zero API cost.

But don't be fooled by the "palace": an independent arXiv paper (2604.21284) found the palace metaphor contributes ≈0 to retrieval performance (it's just ChromaDB metadata filtering). The real contributions are verbatim storage, zero-LLM writes, and minimal wake-up cost (~170 tokens). The paper calls it "a real architectural insight, over-marketed."

Maturity: ~48k stars (growing fast), MIT, Python 3.9+, ~300MB disk. Note: Mem0's April 2026 algorithm has closed the gap to 93.4%.

4. Hermes Agent: The Intern Who Writes Its Own SOPs

Hermes's distinction isn't a smarter model—it's distilling reusable skills from experience at runtime. After roughly every 15 tool calls, the agent automatically reviews and generates skill files (stored in ~/.hermes/skills/), which it calls and improves on similar future tasks.

Three-layer memory architecture:

• Factual memory (MEMORY.md)
• Procedural memory (self-generated files in skills/)
• User profile (USER.md, generated via Honcho)

Notable engineering decisions: six terminal backends (Docker/SSH/Daytona/Singularity/Modal/local)—Modal's serverless sleep/wake means a dormant agent costs nothing until invoked—and six messaging platforms (Telegram/Discord/Slack/WhatsApp/Signal/Email). Compatible with OpenClaw migration.

Maturity: by Nous Research, MIT, Python 3.9+, supports 200+ models.

5. OpenScreen: Native-Grade Screen Recording Post-Production with Web Tech

OpenScreen builds a complete recording post-production pipeline—capture, editing, zoom animations, motion blur, background replacement, annotations, speed control, multi-platform export—with Electron + PixiJS + WebCodecs.

Screen Studio ($29/month) proved "auto-follow mouse zoom" has market value but is closed-source. OpenScreen's bet: implement 80% of core features with a web stack, free and open.

Technical stack:

• Recording: MediaRecorder API + WebAudio mixing
• Rendering: PixiJS (WebGL 2D engine) as a full video compositor
• Export: WebCodecs VideoEncoder + mediabunny MP4 muxer
• Timeline: dnd-timeline for drag-and-drop editing

Key limitation: WebCodecs encoding runs at ~1/7 real-time (about 5fps for 2K content)—one minute of footage takes seven minutes to export. WebCodecs is a software encoder with no hardware acceleration access.

Maturity: MIT, a very new project, early prototype.

Cross-Comparison

Technical Routes

| Dimension | OMX | Edge Gallery | MemPalace | Hermes | OpenScreen | |-----------|-----|-------------|-----------|--------|------------| | Core engine | Codex CLI + Node.js | LiteRT + Kotlin | Python + ChromaDB | Python + LLM APIs | Electron + PixiJS | | Persistent state | .omx/ filesystem | Local SQLite | Local files + SQLite KG | ~/.hermes/ filesystem | Project files | | Parallelism | tmux + multi-worktree | Async concurrent execution | None (serial retrieval) | Sub-agents | None (serial rendering) | | Cost model | Codex API only | Fully free (on-device) | Zero API cost (local embeddings) | LLM API costs | Fully free | | Ecosystem | npm + Codex hooks | Google AI Edge | MCP server (29 tools) | agentskills.io + OpenClaw | Standalone app |

Maturity

| Project | Code Quality | Docs | Community | Stability | Production Readiness | |---------|-------------|------|-----------|-----------|----------------------| | OMX | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | Individuals/small teams | | Edge Gallery | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | Experimental Beta | | MemPalace | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | Production-ready | | Hermes Agent | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | Exploratory use | | OpenScreen | ⭐⭐⭐ | ⭐⭐ | ⭐⭐ | ⭐⭐ | Early prototype |

Trend Insights: Three Shifts Underway

From "Model-Centric" to "Orchestration-Centric"

OMX, Hermes Agent, and similar projects point to one trend: the bottleneck is no longer model intelligence but how multi-agent systems are orchestrated, how state is managed, and how long-horizon tasks avoid collapse. The common theme: state continuity matters more than reasoning capability.

From "Cloud-First" to "On-Device + Local" Dual Track

Edge Gallery (on-device inference) and MemPalace (local-first storage) prove that not all AI needs the network: phones can run 7B+ parameter models locally, and local storage can achieve top retrieval performance at zero API cost.

From "Consumer Tools" to "Production Infrastructure"

OpenScreen's web-stack video editing and OMX's conversational-to-pipeline transformation both show that "commodity technologies" (Web/Node.js/Python) are eating into traditional professional software. When the performance gap shrinks to acceptable levels, development speed and ecosystem richness win.

A Feynman-Style Reality Check

A Feynman-style question: "Have you honestly said what you built versus what you borrowed?"

• OMX: the orchestration layer is original; Codex's reasoning belongs to OpenAI; tmux parallelism is clever engineering, not a new algorithm.
• Edge Gallery: LiteRT and Gemma are Google's; the project is integration, not underlying innovation—though packaging scattered tech into a usable product is itself engineering value.
• MemPalace: the verbatim philosophy is a real insight, but the 96.6% benchmark mostly credits all-MiniLM-L6-v2 + ChromaDB, not the palace metaphor. To its credit, MemPalace v3.3.0 revised its headline from 100% to 98.4%.
• Hermes: the learning loop is appealing, but the "review every 15 tool calls" threshold is a heuristic, not an optimum; skill self-improvement needs stricter validation.
• OpenScreen: the web stack has real costs—software encoding speed is a hard constraint, a trade-off, not a silver bullet.

The takeaway: the most honest projects are those that proactively correct their own headlines. Genuine value comes from those who acknowledge their boundaries.

Conclusion

None of these five projects calls itself "revolutionary"—they are all pragmatic engineering choices. Their value lies not in disruption but in making AI go from "try it out" to "use it daily": lowering barriers, improving reliability, controlling costs, preserving privacy.

The most noteworthy aspect of the 2026 open-source AI ecosystem is not the parameter race, but the engineering, orchestration, and infrastructure work being built around the models.