Walk into any 2026 developer’s terminal and you’ll find at least one AI coding agent already configured. The space moved faster than most people tracking it could follow: terminal-native agents have overtaken IDE plugins for serious engineering work, open-source caught up to commercial tools feature-for-feature (mostly), the Model Context Protocol is the de facto standard for tool and data integration, and the BYOK + self-hosting path is finally a real first-class option. This roundup is the index article of the rest of the series — it maps the landscape, names the names, and tells you which agent to pick for which job.
The five categories that matter in 2026
The market looks like one undifferentiated blur of “AI helpers”, but the actual product categories have settled:
1. Terminal-native agents. Live in the shell, read and edit the repo, run tests, commit. They compose with pipes, scripts, and CI rather than wrapping an editor. Claude Code, Aider, Goose, OpenCode, Hermes Agent, Codex CLI, Junie CLI.
2. IDE-integrated agents. Either VS Code forks rebuilt around AI (Cursor, Windsurf, Trae, Zed) or extensions that turn the existing editor into an agent surface (GitHub Copilot, Cline, Continue, Cody, Roo Code, Tabnine, Qodo, JetBrains Junie in IDE).
3. Self-hosted open-source assistants. Inference and model server run on the developer’s own hardware or a private cluster, no code leaving the network — the privacy / compliance path. Tabby, Continue + Ollama, OpenHands, Aider, Goose, OpenCode, Cline, Hermes Agent.
4. Autonomous / async cloud engineers. Take a ticket or goal, return a finished PR with minimal supervision — headless, cloud-hosted, often metered per task. Devin, OpenHands Cloud, Codex cloud, GitHub Copilot coding agent, Factory, Replit Agent.
5. No-code / prompt-to-app builders. Prompt in, deployed app out — target non-developers and prototyping. Bolt.new, Lovable, v0 by Vercel, Replit, Google AI Studio.
This article focuses on the first two, with a passing mention of the third. Categories 4 and 5 are different products for different buyers.
The 2026 trends that shape the comparison
Agentic loops are the default. Plan → edit → test → iterate → commit. Pure chat assistants are being deprecated inside the same vendors. Claude Code defaults to plan-first; OpenCode has /plan; Cline has Plan+Act toggles; Goose’s defaults lean toward Act. If your agent of choice still insists on pure turn-by-turn chat, it’s lagging.
Model Context Protocol is everywhere. Anthropic-reported 10,000+ public MCP servers by early 2026; supported across Claude Code, ChatGPT, Cursor, Windsurf, Copilot, Junie, VS Code, Goose, Cline, and the rest. Tool and data access is now a marketplace, not a per-tool integration.
Terminal-native has won the experienced-developer vote. Developers who already live in the shell don’t want to switch to a new editor to get AI; they want pipes, scripts, and CI to compose well. Every “agent that lives in the editor” has at least one terminal sibling trying to meet the same need from the other side.
Open-source caught up. Aider, Continue, OpenHands, Tabby, Cline, Goose, and OpenCode cover most of what Cursor / Copilot / Claude Code offer, often with BYOK and self-hosting. The gap is no longer feature set — it’s polish, integration depth, and the MCP marketplace.
Subscription fatigue and credit metering. Flat $20/mo tiers are converging across Cursor, Windsurf, GitHub Copilot, Claude, and ChatGPT. Heavy agentic work moves to per-token or per-ACU pricing. Open-weight local inference is a real option again.
Comparison axes that actually matter
For terminal-native agents specifically, the columns that drive 90% of decisions:
- License — Apache-2.0 / MIT for OSS, BSL or proprietary for commercial. Compliance teams care about the first group; everyone else is fine with either as long as price works.
- Provider model — single-vendor (Anthropic for Claude Code, OpenAI for Codex) vs multi-provider / BYOK. The multi-provider path is a hedge against model-provider price hikes.
- Self-hostable — Yes (Tabby, OpenHands, Aider, Goose, OpenCode, Cline) / partial / No (Claude Code, Junie, Cursor, Windsurf).
- Pricing — Free OSS + BYOK API (Aider, Continue, Cline, Goose, OpenCode, OpenHands, Tabby) / Free + ~$20/mo Pro (Claude Code, Cursor, Windsurf, Copilot, ChatGPT, Cody, Junie, Bolt.new, Lovable, v0, Replit) / ~$40-100/mo heavy tier / ~$500/mo enterprise/ACU (Devin).
- Plan-first default — Native (Claude Code, Junie Plan Mode, Windsurf Cascade, Cline, Roo Code, Goose, OpenCode) / optional / minimal.
- Multi-file edit confidence — long-running agent (Claude Code, Codex CLI, Goose, OpenCode, Aider, Amp) > in-editor agent (Cline, Continue, Roo Code, Junie) > autocomplete (Tabnine, Cody basic).
The seven agents this series covers
| Agent | Category | License | Provider | Plan-first | Best for |
|---|---|---|---|---|---|
| Hermes Agent | Terminal | MIT | BYOK multi | No (Act default) | Self-hosted, multi-platform gateway |
| Hermes Desktop | Terminal | MIT | (inherits) | (inherits) | Hermes users wanting a GUI |
| Claude Code | Terminal | Proprietary | Anthropic | Yes | Solo power user, plan-first discipline |
| OpenCode | Terminal | MIT | BYOK multi | Yes (/plan) |
Provider-agnostic, TUI-first |
| Aider | Terminal | Apache-2.0 | BYOK multi | Optional | Auto-commit, fast diff-and-commit loop |
| Cline | IDE | Apache-2.0 | BYOK multi | Yes (Plan+Act) | VS Code users, diff-in-place UX |
| Goose | Terminal | Apache-2.0 | BYOK multi | No (Act default) | Static-binary deploy, AAIF governance |
Each gets its own article in this series with hands-on install, day-one workflow, and the configuration patterns that earn real estate.
Picking one in 2026
- You do most of your work in a terminal, you want one subscription that just works, you’re willing to read plans instead of accepting diffs blindly → Claude Code. Highest leverage per dollar for solo engineers.
- You want provider flexibility and don’t want to be locked to Anthropic → OpenCode if you’re a TUI person, Cline if you want to stay in VS Code.
- You want MIT/Apache source code you can audit, fork, or self-host → Aider for fast commits, Goose for a single-binary deploy, Hermes Agent for multi-platform gateway + memory.
- You want a strict Plan+Act toggle in the editor with diff previews on every step → Cline. The most disciplined UX in the field.
- You already use Ollama locally and just want a sane shell → OpenCode with the local provider adapter, or Hermes Agent.
What I deliberately didn’t pick
Cursor and Windsurf. Both are excellent in 2026, but they’re a switch away from VS Code, not a sidebar. Different ergonomic, different conversation. If you’re an “always in the IDE” person, those are the picks; this series is for terminal and editor-extension agents.
Devin and autonomous cloud engineers. Different product for different buyer. If the question is “can a thing take a Jira ticket and ship a PR by morning”, those are where to look; this roundup is about the agent sitting next to you.
Bolt.new / Lovable / v0. Not engineering tools, prototype tools. They’re great at what they do; they don’t show up in a Linux developer’s daily workflow.
Closing thought
The 2026 landscape isn’t “which agent wins” — it’s “which agent wins for the kind of work you actually do“. Pick the one that matches your shell discipline (terminal-native), your editor (IDE-extensions), or your governance model (open-source/AAIF), and commit to reading plans instead of accepting diffs. The rest is configuration.
Comments