Table of Contents

• The Comparison Table
• opencode: The One That Grew the Most and Why
• The VS Code Extensions: Cline and Kilo Code

  • Cline
  • Kilo Code
• The Premium Terminal Agents: Crush and Droid

  • Crush
  • Droid (Factory)
• The Cost Factor: Why GLM Shows Up in Every Testimonial
• Recommendations by Profile
• The Setup People Who've Been at It a While Actually Use
• Sources

The Comparison Table

Tool	Type	License / Model	Entry Price	Excels At	For Who
opencode	Terminal Agent (TUI)	Open Source (SST)	Free + BYOK, or Zen pay-as-you-go	75+ model providers, Plan/Build modes, LSP support	Devs who live in the terminal and don't want vendor lock-in
Claude Code	Terminal Agent + IDE	Proprietary (Anthropic)	$20/month (Pro plan)	Complex refactors, multi-file coherence	Who prioritizes code quality over cost
Codex CLI	Terminal Agent + IDE	Proprietary (OpenAI)	Included in ChatGPT Plus ($20/month)	PR reviews, pipeline tasks, CI	Teams already paying for ChatGPT
Cline	VS Code Extension	Open Source (Apache 2.0)	Free + BYOK	Plan-then-execute architecture with step-by-step approval	Devs who want total control over every edit
Kilo Code	VS Code / JetBrains Extension	Open Source (Cline/Roo fork)	Free + BYOK or credits	Orchestrator mode, Memory Bank, inline autocompletion	Teams wanting a supported stack with centralized billing
Crush	Terminal Agent (TUI)	Open Source (Charm)	Free + BYOK	Best terminal experience, LSP + MCP	Who values a beautiful TUI and project-based sessions
Droid (Factory)	CLI + desktop + headless	Proprietary	$20/month (Pro)	droid exec for CI/CD, intelligent model routing	Enterprise teams automating pipelines
Cursor	Complete IDE (VS Code fork)	Proprietary	$20/month (Pro)	The most balanced and mature package	The default if you want a single tool that does everything
Antigravity	Agentic IDE	Proprietary (Google)	Free tier; Pro $20/month	Manager View with parallel agents, Chrome integration	Who works with Gemini and needs to orchestrate multiple agents
Zed	Native Editor	Open Source	Free; Pro $10/month	Brutal performance (120fps, ms startup)	Who hates Electron's slowness and brings their own agent

---

opencode: The One That Grew the Most and Why

OpenCode is a terminal-native code agent, open source, built by the team behind SST. By April 2026, it had accumulated 147,000 GitHub stars and about 6.5 million monthly developers, growing nearly 4.5 times faster than Claude Code over the same period.

What makes it different:

Supports 75+ model providers out of the box. You're not tied to a single lab. You can run Claude for heavy refactoring, a cheap model for boilerplate, and something local for sensitive work, without switching tools.

Plan and Build modes. Plan explores, reads, and proposes without touching anything. Build executes. That separation is the difference between "the agent understood the problem" and "the agent rewrote three files it shouldn't have."

Real LSP support. It understands your code with the same intelligence as your IDE, not by guessing from plain text.

Zero storage of your code. A relevant point if you work with clients who have confidentiality clauses.

For models there are two paths: BYOK (you bring your API key and your provider charges you) or opencode Zen, a gateway with models already tested and benchmarked for agents, pay-as-you-go with no markup on requests.

💡 Get $5 in credit for your usage limits If you sign up with this OpenCode Go link, you get $5 in credit applied directly to your usage limits. It practically covers your first month of testing before you spend a cent.

👉 Subscribe with 10% off on Z.ai here — the discount applies only to your first payment and only if you've never had a paid subscription on the platform.

How I use it: terminal in a split pane next to my editor. I describe the change in natural language, review the diff, accept or reject. Plan mode first when I don't know the module well.

One piece of context that helps understand the ecosystem: opencode and Crush were born from the same split in mid-2025. Charm hired the original creator and renamed their branch Crush; the collaborators on the SST side kept the name OpenCode. That's why they look so alike, and why both are good.

---

The VS Code Extensions: Cline and Kilo Code

If your workflow lives inside the editor rather than the terminal, these are the two that matter.

Cline

Cline is the most established of the open source lineage, with an Apache 2.0 license. Its architecture is plan-then-execute: it plans the complete strategy before writing any code, lets you review the plan, and then executes step by step asking for approval at every stage.

Important: it's an agentic tool, not an autocompletion engine. It doesn't give you inline completions Copilot-style. If you want both, you need to combine it.

"Cline works with a huge variety of model providers, and the GLM Coding Initiative stands out as one of our smoothest integrations. With developer-friendly pricing, generous quotas, and latency that fits perfectly into real agentic workflows, its value proposition is hard to beat." — the Cline team

Use it if: you want a lean agent, a permissive license, broad BYOK, local inference, and programmable workflows via CLI/SDK.

Kilo Code

Kilo Code forked both Cline and Roo Code and added what both were missing: orchestrator mode (automatically routes tasks to specialist modes), inline autocompletion, and a Memory Bank to persist project context.

Important context note for 2026: Roo Code shut down on May 15, 2026 and its repository was archived. If you're coming from there, Kilo maintains the migration path. It has also expanded the stack to JetBrains, CLI, Cloud Agents, subagents, and Agent Manager.

"Developers don't just care about raw 'IQ': they care about high-efficiency output without worrying about costs or restrictions. The GLM Coding Initiative tackles that pain head-on. Combined with Kilo Code, it lets you forget the anxiety and focus on the heavy lifting of real development." — the Kilo Code team

Use it if: you're a team that wants a single supported stack, a large catalog of hosted models, a native experience in VS Code and JetBrains, and centralized billing and governance.

---

The Premium Terminal Agents: Crush and Droid

Crush

Crush is Charm's entry into the agent race. Written in Go, with LSP-aware context, MCP support, per-project session management, and multi-model support (Anthropic, OpenAI, Gemini, Groq, OpenRouter, and local endpoints). More than 25,000 stars since May 2025.

Its honest differentiator is the terminal experience. Charm has been making the best TUIs in the ecosystem for years, and it shows.

"As one of today's leading model families, the GLM series — with its open source roots and aggressive pricing — shows the team's commitment to competing at the highest level. We use GLM to design critical components of Crush and the experience has been exceptional. We recommend it to any team serious about building products on top of LLMs." — the Crush team

Use it if: you live in the terminal, value visual polish, and want a fast Go agent with MCP.

Droid (Factory)

Factory plays in a different league: enterprise. Droid deploys across three surfaces — CLI (droid), desktop app, and headless mode (droid exec) designed for CI/CD and automation.

Pricing starts at $20/month (Pro), with Plus at $100 and Max at $200. There is no free tier, which is a real obstacle when evaluating it against Cursor or Claude Code. In exchange: MCP support, multi-agent, deep Git integration, multi-file editing, automatic test execution, and access to frontier models. They raised a $150M Series C at a $1.5B valuation in April 2026, with Nvidia, Adobe, EY, and Adyen as customers.

"GLM-5 delivers frontier intelligence at a fraction of the cost. It thinks through system design before coding, and handles complex refactors with ease. Built for resilience: it maintains perfect coherence in multi-step workflows and generates clean documentation with professional diagrams." — the Factory team

Use it if: your bottleneck isn't writing code but automating the entire pipeline, and you have an enterprise budget.

---

The Cost Factor: Why GLM Shows Up in Every Testimonial

There's an obvious pattern in the quotes above, and it's no coincidence. The defining feature of the open source agent market in 2026 is BYOK: zero subscription, but API costs that scale with usage. A heavy user working with frontier models spends between $50 and $200 a month on tokens alone.

That's where alternative model subscription plans come in. The GLM Coding Plan has tiers at $18 (Lite), $80 (Pro), and $168 (Max) per month, with 30% off annual billing, and it works inside Claude Code, Cline, Kilo Code, and 20+ other clients.

The practical takeaway: the cost of the agent and the cost of the model are separate decisions. You can use a free open source tool with a cheap model plan, or pay $20 for a proprietary tool with its model included. Run the numbers with your actual volume before deciding.

---

Recommendations by Profile

If you're a solo dev and want to start today without spending: opencode with BYOK, or Zed as your editor + whichever agent you prefer. Zero fixed cost, total control.

If you want a single tool that does everything well: Cursor. Still the default for overall balance, model selection, and maturity of agentic features.

If your job is refactoring complex legacy code: Claude Code. Its multi-file coherence justifies the price.

If you already pay for ChatGPT Plus: Codex CLI is effectively free for you. Excellent for automated PR reviews.

If you work inside VS Code and want granular control: Cline. If you also want autocompletion and orchestration, Kilo Code.

If you need to run several agents in parallel: Antigravity with its Manager View, especially if Gemini is your preferred model — it's the cheapest way to use it in an IDE.

If your problem is the pipeline, not the code: Droid with droid exec.

If the TUI matters to you as much as the result: Crush.

---

The Setup People Who've Been at It a While Actually Use

No senior dev in 2026 uses a single tool. The real pattern is a deliberate split:

• A powerful agent for the hard stuff — refactors, debugging weird bugs, architecture decisions.
• A cheap, fast agent for the repetitive stuff — boilerplate, tests, docs, mechanical migrations.
• An editor that stays out of the way — fast to start and doesn't eat your RAM.

My concrete recommendation for building your stack: start with opencode + an economical model plan, measure your actual spending for a month, and only then decide whether it's worth paying $20 for a proprietary tool. It's much easier to move up a tier than to untangle yourself from vendor lock-in.

---

Sources

• OpenCode — official site and Zen documentation
• Best Terminal AI Coding Agents in 2026 — amux
• Best Open-Source AI Coding Agents 2026: Cline vs Roo vs Kilo vs Aider
• Roo Code vs Cline vs Kilo Code — Kilo
• charmbracelet/crush on GitHub and Crush, Welcome Home — Charm
• Factory AI Review 2026 — The AI Agent Index
• Cursor vs Windsurf vs Antigravity — andrew.ooo
• AI Code Editor Comparison 2026 — devgent
• Claude Code vs Codex CLI vs Gemini CLI 2026 — RockB
• GLM Coding Plan Guide 2026