The open-source coding CLI landscape, mapped
Two years ago the open-source coding agent scene was a hundred GitHub repos with "agent" in the name, most of them demos that died a week after the launch post. By mid-2026 it has consolidated into roughly six projects you can bet a workflow on. They are not interchangeable. Each one made a different structural bet, and picking the wrong one for your situation costs real time.
I covered the vendor side (Claude Code, Codex, Gemini CLI) in my 2026 tools overview. This piece maps the other side of the fence: the tools whose roadmap no platform strategy depends on.
The six that matter
| Tool | The bet | Reach for it when |
|---|---|---|
| Aider | Git-native edit loop with a ranked repo map | You want surgical edits and clean history |
| OpenCode | Polished TUI, any provider, client/server core | You live in the terminal and switch models |
| OpenHands | Sandboxed autonomy with research rigor | You hand off whole tasks, not single edits |
| Goose | MCP-first extensions plus shareable recipes | Your workflows reach beyond the repo |
| Crush | Terminal craft from the Charm ecosystem | You want the slickest multi-model TUI |
| Qwen Code | Gemini CLI fork tuned for Qwen models | You want a capable agent on a free tier |
Two lineage notes, because they explain the shape of the field. Crush grew out of an earlier project in the same niche and joined the Charm ecosystem, which is why its terminal rendering embarrasses most of this table. And Qwen Code is a literal fork of Google's Gemini CLI, retargeted at Qwen's coding models. That is open licensing doing exactly what the label promises: the forks are load-bearing.
Don't read the table as six flavors of the same thing. Aider is a precision hand tool. OpenHands is a workbench with a safety cage around it. Goose barely thinks of itself as a coding tool at all. "Open-source coding CLI" is a category the way "vehicle" is a category.
What open actually buys
Strip the ideology away and three concrete things remain:
- Model choice. Point the same tool at a frontier Claude today, a GPT-5-class model tomorrow, and a quantized local model on a plane. Route cheap work to cheap models. Vendor CLIs are only now loosening their grip here; open CLIs never had a grip to loosen.
- Auditability. When the agent does something strange, you can read the prompt assembly, the context selection, the tool schemas. As agents get more autonomous, "I can read the harness" shifts from a nicety to a requirement.
- No lock-in. A pricing change, an acquisition, a sunset notice. None of it strands your workflow. The Qwen Code fork is the existence proof that the escape hatch actually opens.
There is a fourth, quieter benefit: these codebases are the best textbooks we have on agent design. If you want to understand the control loop instead of just riding it, build a toy agent yourself, then read how Aider or OpenHands solved the parts you fudged.
What it costs
Open means the integration work moves from the vendor's payroll to yours.
- Assembly required. You choose the provider, the model, the keys, the permission posture, the context strategy. Defaults exist; none of them were tuned for your stack.
- The tuning gap. Vendor agents co-evolve with their models: harness and model are, by all appearances, trained toward each other. An open harness has to stay generic, and the same model often performs slightly worse in it, especially at recovering from weird states.
- Support is a GitHub issue. No SLA, no escalation path. Irrelevant for a side project; a real line item when you're standardizing a team on one of these.
- Licenses vary. Most of the six are Apache-2.0 or MIT, but not all, and terms have been known to shift. Read the license before anything load-bearing goes on top.
How to pick
Match the tool to the unit of work you hand it. Surgical edits with defensible git history: Aider. A terminal home that survives model churn: OpenCode or Crush. Whole tasks executed in isolation: OpenHands. Automation that leaves the repo (tickets, docs, internal systems): Goose. Maximum capability at zero spend: Qwen Code.
Whether you should be on this side of the fence at all is the prior question: vendor CLIs buy you tuning, support, and momentum with someone else's roadmap attached, and I weigh that trade properly in open CLI vs vendor. But if you already know you want open, 2026 is a good year for it. This is no longer the compromise aisle. It's six real tools with six real opinions, and at least one of them matches how you work.