惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Microsoft Azure Blog
Microsoft Azure Blog
aimingoo的专栏
aimingoo的专栏
F
Fortinet All Blogs
Blog — PlanetScale
Blog — PlanetScale
GbyAI
GbyAI
MongoDB | Blog
MongoDB | Blog
月光博客
月光博客
The Cloudflare Blog
量子位
T
Tailwind CSS Blog
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
B
Blog
MyScale Blog
MyScale Blog
T
The Blog of Author Tim Ferriss
The GitHub Blog
The GitHub Blog
G
Google Developers Blog
D
DataBreaches.Net
V
Visual Studio Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
Last Week in AI
Last Week in AI
U
Unit 42
博客园 - 聂微东
有赞技术团队
有赞技术团队
A
About on SuperTechFans

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
OpenCode Hit 140K Stars. Why Terminal Agents Won 2026.
jidonglab · 2026-04-24 · via DEV Community

140,000 stars. 850 contributors. 11,000 commits. 6.5 million developers using it every month. Zero IDE integration.

OpenCode is a terminal coding agent. It runs in your shell. It has no VS Code extension, no JetBrains plugin, no web UI. When it launched in March 2025 as a Go-based alternative to Aider and Cline, the conventional wisdom was that terminal-only was a deliberate niche — the kind of thing a few vim users would love and the rest of the market would ignore.

The rest of the market did not ignore it. Between January and April 2026, OpenCode crossed Cline, crossed OpenHands, and closed the gap on Aider despite Aider's two-year head start. I've been using it daily since February. Here's why the terminal won and what OpenCode's architecture got right that the IDE-bound tools missed.

The definition: what a terminal coding agent actually is

A terminal coding agent is a command-line process that reads your codebase from disk, talks to an LLM, writes diffs to files, and commits changes via git. The interface is whatever your terminal supports — text, a TUI, keybindings. There is no editor integration because the editor is wherever you want it to be.

That sentence contains the entire argument for terminal agents. The editor is wherever you want it to be. It can be Neovim on a remote dev box, VS Code on a laptop, Helix in a tmux session, or no editor at all if you're doing a batch migration. The agent doesn't care. It operates on files, not on buffers.

Your editor (anything) ←→ Files on disk
                              ↑
                      OpenCode (terminal)
                              ↓
                    LLM (75 supported models)

Enter fullscreen mode Exit fullscreen mode

This decoupling is the thing. It's also the thing that Cline and Cursor explicitly rejected. Both bet that deep IDE integration — selection-aware context, inline diffs, click-to-apply — would be the defining UX. They weren't wrong about the UX. They were wrong about the market.

Why IDE integration stopped being a moat

I watched this happen in real time. In January 2026, Cline was the highest-starred AI coding agent with deep IDE integration. By April, it had been passed by three terminal-first tools. The stars were the symptom; the cause was workflow.

Three workflow shifts happened in the six months after late 2025:

First, remote dev environments became table stakes. GitHub Codespaces, Gitpod, and self-hosted dev containers became how serious teams worked. Every engineer I know who ships to production now SSHs into a box they didn't provision, edits files with whatever editor is installed, and commits from a terminal. An IDE-bound agent requires you to also forward your IDE to the remote box, which most people don't bother doing. A terminal agent is already there.

Second, Claude Code normalized the terminal for AI coding. Anthropic's own tool shipped as a CLI. Millions of developers who had been skeptical of terminal workflows used one every day because Anthropic's did. OpenCode rode that wave directly.

Third, multi-machine workflows got normal. You write code on your laptop, deploy to a cloud box, run agents on a local workstation for heavy jobs. An IDE extension has to run where the IDE runs, which is one place. A terminal agent runs anywhere you have a shell, which is everywhere.

The result wasn't that IDE extensions died — they didn't. It was that they stopped being the default answer. A developer asking "what coding agent should I use" in January 2026 got pointed to Cline. The same question in April got pointed to OpenCode, Aider, or Claude Code, depending on budget and taste.

The architecture decision that made OpenCode fast

I ran OpenCode against Aider and Cline on the same task for a week. The task was mid-complexity: refactor a Next.js app to use Server Actions across 34 files. Aider and Cline averaged 18-24 seconds per file. OpenCode averaged 6.

The difference was the language. OpenCode is written in Go. Aider is Python, Cline is TypeScript running in the VS Code extension host. For a tool that spends its time reading files, parsing diffs, and piping text to an LLM, Go's concurrency primitives and fast startup matter more than they should. OpenCode opens the repo, loads a file tree, and is ready to accept a prompt in under 150ms. Cline, running inside VS Code's extension host, takes 1.2-2 seconds to become responsive because it has to wait for the TypeScript runtime and the extension API.

// opencode/session/session.go — the core session loop, simplified
func (s *Session) Run(ctx context.Context, prompt string) error {
    plan := s.planner.Plan(prompt)
    for _, step := range plan.Steps {
        go s.execute(ctx, step)
    }
    return s.wait(ctx)
}

Enter fullscreen mode Exit fullscreen mode

Each step runs in its own goroutine. Reading files, calling the LLM, writing diffs — they happen in parallel when they can, which is often. Aider's equivalent loop is synchronous Python. Cline's is callback-driven TypeScript that runs inside VS Code's single extension host thread. For a task that touches 34 files, that throughput difference compounds.

The second architectural call was model routing. OpenCode supports 75 models via OpenRouter, direct Anthropic, direct OpenAI, Ollama for local models, and a custom-endpoint mode. You can route planning to Claude Opus 4.7 and execution to Claude Haiku 4.5 via a single config flag. On that Next.js refactor, I saved roughly 60% on token cost by routing the mechanical edits to Haiku while keeping the planning on Opus.

# opencode.config.yaml
models:
  planner: anthropic/claude-opus-4-7
  executor: anthropic/claude-haiku-4-5
  fallback: openrouter/qwen-2.5-coder-32b

Enter fullscreen mode Exit fullscreen mode

This is a pattern I also wrote about in

Post not found or has been removed.

— route the expensive model to the expensive problem, the cheap model to the mechanical work. OpenCode bakes it into the config.

The two UX decisions that actually matter

I've written a lot of agents. What I underestimated about OpenCode was how much of its appeal came from two specific UX choices that look small on paper.

The first is dual-mode agents. OpenCode has a Build mode and a Plan mode. Build mode writes diffs immediately. Plan mode produces a written plan and waits for your approval before touching a file. You switch between them with a keystroke. I did not think this mattered. It matters enormously. The friction of approving a plan before execution is exactly low enough that you do it for anything bigger than a one-file change, and the errors it catches are exactly the kind of errors that would otherwise cost you 40 minutes of git reset --hard.

The second is multi-session. You can run multiple OpenCode sessions in parallel on different branches of the same repo. Each session has its own model config, its own conversation state, its own plan. I typically run two: one on the branch I'm actively reviewing, one on a long-running migration that needs attention every few hours. Aider doesn't support this cleanly. Cline doesn't support it at all because it's tied to a single VS Code window.

These sound like small wins. In practice they eliminate two of the three biggest sources of friction in agent-based coding: "did I just blow up a file" and "do I have to context-switch my agent every time I context-switch my task."

Where Aider still wins and where OpenCode doesn't try

OpenCode is not strictly better than Aider. Aider is older, more stable, and has the best git integration I've seen in any coding agent. If your workflow is "make a small edit, review the diff, commit immediately, repeat," Aider is tighter. Its --commit flag does exactly what you want without you needing to think about it.

OpenCode is better for larger refactors, multi-file edits, and anything involving model routing. It's also better if you want to hand off to a teammate — the session state is serializable and portable, so you can pass an in-progress agent session to a colleague along with the branch.

Neither of them is Cline. Cline still wins for a specific flow: you're editing inside VS Code, you want to see inline suggestions, you want selection-aware context. For that workflow Cline is unmatched. It's just not the workflow that most developers optimized for in 2026.

What to actually do with this

If you're still on an IDE-bound coding agent, try a week with a terminal one. Not as a lifestyle change — as a benchmark. The terminal tools got fast enough that the context you lose from not having IDE integration is smaller than the speed you gain from cold-starting an agent in 150ms instead of 1.5 seconds.

If you're building a coding agent, the lesson from OpenCode isn't "write it in Go." The lesson is that the bottleneck is no longer model quality or prompt engineering. It's startup time, routing flexibility, and the ability to survive in the workflows developers actually use, which are increasingly remote, multi-machine, and terminal-primary. If your agent doesn't run on an SSH-only dev box, you're losing the next generation of users.

The OpenCode repo:

OpenCode logo

The open source AI coding agent.

Discord npm Build status

English | 简体中文 | 繁體中文 | 한국어 | Deutsch | Español | Français | Italiano | Dansk | 日本語 | Polski | Русский | Bosanski | العربية | Norsk | Português (Brasil) | ไทย | Türkçe | Українська | বাংলা | Ελληνικά | Tiếng Việt

OpenCode Terminal UI


Installation

# YOLO
curl -fsSL https://opencode.ai/install | bash
# Package managers
npm i -g opencode-ai@latest        # or bun/pnpm/yarn
scoop install opencode             # Windows
choco install opencode             # Windows
brew install anomalyco/tap/opencode # macOS and Linux (recommended, always up to date)
brew install opencode              # macOS and Linux (official brew formula, updated less)
sudo pacman -S opencode            # Arch Linux (Stable)
paru -S opencode-bin               # Arch Linux (Latest from AUR)
mise use -g opencode               # Any OS
nix run nixpkgs#opencode           # or github:anomalyco/opencode for latest dev branch

Enter fullscreen mode Exit fullscreen mode

Tip

Remove versions older than 0.1.x before installing.

Desktop App (BETA)

OpenCode is…



This is the final part of the series. Part 1 covered the Skills paradigm and how Karpathy's observations became a loadable format. Part 2 covered OpenClaw and why local-first agents beat cloud ones. All three projects tell the same underlying story — 2026 was the year the agent stopped being a product and started being a primitive.

The best coding agent is the one that boots before you finish typing the first word of your prompt.

For anyone who's made the switch from IDE to terminal agents recently — what was the specific task that convinced you? I'm collecting these because the narrative is still forming and I think the tipping point for most people was a single workflow, not a gradual preference shift.


Sources: