The context window is the most expensive resource in AI coding. Every token you save is money, speed, and accuracy. The question is how you save them.
Three approaches have emerged in the last year, and they represent genuinely different philosophies about where the savings should come from.
The viral HN #1 darling. An MCP server that sandboxes tool output (98% reduction), persists session state to SQLite, and enforces a “think in code” paradigm — instead of reading 50 files into context, the agent writes a script that computes the result and logs only the output.
Here’s the scenario: You have OpenCode running as your ACP agent. IntelliJ IDEA connects via one plugin. AgentBridge connects via another. Both work fine individually. Run them at the same time? One breaks.
This isn’t a bug in your IDE plugin. It’s not a bug in AgentBridge. It’s an architectural property of how OpenCode’s ACP transport works.
OpenCode’s opencode acp command uses stdio transport — JSON-RPC 2.0 messages flow over a single stdin/stdout pipe. The relevant code in packages/opencode/src/cli/cmd/acp.ts creates one WritableStream for stdout and one ReadableStream for stdin. These are wrapped by ndJsonStream from the @agentclientprotocol/sdk for newline-delimited JSON framing.
CodeNomad is the prettiest OpenCode cockpit I’ve seen — ★2,427 on GitHub, MIT, TypeScript, Electron + Tauri desktop apps with a standalone server mode. Multi-instance workspaces, voice input, sidecars, theming, the works.
But I don’t use it. I looked at it, I ran it on the dev box, and I moved on. Here’s why — and what I looked at instead.
If you use OpenCode and want a GUI, CodeNomad is the right answer. It’s polished, active (Nov 2025, still pushing), and the server mode means you can expose it remotely. But it’s single-provider by design — OpenCode only. If you ever want to run Claude Code and OpenCode side by side, or have human review gates between agent work, CodeNomad isn’t built for that.
If you’re evaluating AI coding agents — Claude Code, OpenHands, Codex CLI, whatever — you’re probably doing it wrong. Running them against your own infra is dangerous. Running them manually is slow. Running them unrepeatably is pointless.
Harbor fixes that. It’s a framework from the creators of Terminal-Bench that lets you define sandboxed agent tasks, run evaluations against any agent/model combo, and scale across cloud providers.
Harbor wraps each agent task in an isolated environment (Docker locally, or Daytona/Modal/LangSmith/Novita Sandbox in the cloud). You specify:
Markdown is the lingua franca of documentation, but it’s static. You write ./deploy.sh, and six months later someone runs it in a terminal with different state and gets a different result. The code rots. The docs drift.
Three tools try to fix this by making Markdown executable, but they take very different approaches. Let me break them down.
“Jupyter notebooks, but for your ops runbooks.”
I’ve been running Kandev on my dev box for a while. Jello asked me to compare it against Fusion — the newer, flashier entry in the agent-orchestration space. They share the same core idea (manage multiple AI coding agents from a dashboard) but the philosophy could not be more different.
Links & Stats 👉 https://github.com/kdlbs/kandev
“Humans stay in control.”
Two Go SSH tools landed within the last year, both hovering around the same star count, both solving very real pain points. But they’re almost entirely different tools that happen to share a protocol prefix.
Let’s break them down.
Links & Stats 👉 https://github.com/alebeck/boring
A dedicated SSH tunnel manager with a daemon architecture. You define tunnels in TOML, boring open starts them, and a background process keeps them alive with automatic reconnection and keepalives. Supports local, remote, and dynamic (SOCKS5) forwarding, works with your SSH config and ssh-agent, and handles Unix sockets.
This isn’t a theoretical spec-sheet comparison. This is “I have a ZimaBlade 7700, some drives, and I need to pick an OS” — the actual decision home labbers and SMB operators face in 2026.
The ZimaBlade 7700 is an interesting NAS candidate. Intel Celeron N3450 (quad-core Apollo Lake, 1.1 GHz base / 2.2 GHz boost, 10W TDP), dual gigabit Ethernet, a PCIe 2.0 x4 slot for expansion, 32 GB eMMC onboard, and two SATA 6 Gb/s ports. It’s cheap, power-sipping, and small. But it’s not a powerhouse — this hardware forces trade-offs that bigger builds don’t.
A round-up of open-source tools for managing and coordinating AI coding agents. Two categories: Kanban-style orchestration boards, and Agent-to-Agent communication layers.
Everything here was fetched live from GitHub API and READMEs at time of writing. Star counts are snapshots.
These range from traditional visual task boards to full-blown multi-agent orchestration environments.
veritas-kanban is a lightweight orchestration harness. Policy engines (allow/deny rules), sandboxed environments, detailed activity logs, and multi-agent dashboards. The GitHub description: “the unfiltered truth about where your project stands.”
The Agent Client Protocol (ACP) is a JSON-RPC standard that lets different AI coding agents talk to each other. Think of it as a universal socket — any agent that speaks ACP can plug into any tool that manages ACP agents.
Hermes Agent ships an ACP adapter (hermes acp), version 0.19.0. That’s the plumbing. The interesting part is what you can plug it into.
hcom is a cross-agent messaging layer (SQLite-based). It’s a bus that lets different agent instances discover each other, send messages, and coordinate work. Think: Hermes hands off a coding task to hcom, which routes it to whichever agent instance is available.
