Terminal coding agents compared
Nine coding agents that run in your terminal, side by side: what each costs, which models it can use, whether it is open source, and what it scored on Terminal-Bench 2.1. Checked October 17, 2026.
| Agent | License | Cost | Models | Local models | Headless | MCP | OS sandbox | Terminal-Bench 2.1 |
|---|---|---|---|---|---|---|---|---|
| AgenC Core | MIT | Free, bring your own key | 25 providers | Yes | agenc -p | Client and server | Opt-in (Seatbelt, bubblewrap, Landlock) | 76.2% for $34, DeepSeek V4.1 Flash (public run) |
| Claude Code | Closed source | Claude Pro plan or higher, or API | Claude only | No | claude -p | Client and server | Seatbelt, bubblewrap (not native Windows) | 83.8% for $553, Fable 5 |
| Codex CLI | Apache-2.0 | ChatGPT plan or API key | OpenAI, custom providers | Yes (--oss) | codex exec | Client | Seatbelt, bubblewrap, Windows | 87.4% for $774, GPT-6 Astra |
| Gemini CLI | Apache-2.0 | Free daily quota, paid plans | Gemini only | Not documented | gemini -p | Client | Seatbelt, Docker, Podman and more | 65.8% for $236, Gemini 3.1 Pro |
| Cursor CLI | Closed source | Cursor subscription | Cursor and third-party models | Not documented | agent -p | Client | Seatbelt, Landlock | 79.3% for $134, Grok 4.5 |
| OpenCode | MIT | Free, bring your own key | 75+ providers | Yes | opencode run | Client | None | Not listed |
| Pi | MIT | Free, key or subscription login | Many providers | Yes | pi -p | Client | None built in | Not listed |
| Goose | Apache-2.0 | Free, bring your own key | 15+ providers | Yes | goose run | Client | None | Not listed |
| Aider | Apache-2.0 | Free, bring your own key | Most providers | Yes | aider -m | No | None | Not listed |
Terminal-Bench 2.1 shows each agent's best entry on the Harbor Hub board, with the model it used and the cost Harbor Hub lists. AgenC's figure is its public run on Harbor Hub, not a board entry. A score depends on the model as much as on the agent, so compare runs that use the same model when you can. Full results
Which one fits
- Lowest cost for a high score: AgenC Core with DeepSeek V4.1 Flash, 76.2% for $34. Every entry on the Harbor Hub board spent at least $134.
- The highest scores: Codex CLI with GPT-6 Astra and Claude Code with Fable 5 lead the board, at $774 and $553 a run.
- Any model, including local ones: AgenC Core, OpenCode, Pi, Goose and Aider all take your own keys and local models. AgenC Core has 25 providers built in. AgenC's providers
- A sandbox for commands: AgenC Core, Claude Code, Codex CLI, Gemini CLI and Cursor CLI can run commands in an OS sandbox.
- A free tier without a key: Gemini CLI gives a daily quota with a Google account.
Sources
- Claude Code license
- Claude Code setup and plans
- Claude Code model gateways
- Codex CLI license
- Codex pricing
- Gemini CLI quotas and pricing
- Cursor CLI headless mode
- OpenCode providers
- OpenCode security
- Pi security
- Goose providers
- Aider models
- Terminal-Bench 2.1 board on Harbor Hub
- AgenC's public Terminal-Bench 2.1 run
Questions
What is the cheapest AI coding agent with a high score?
AgenC Core with DeepSeek V4.1 Flash scored 76.2% on Terminal-Bench 2.1 for $34 in a public run. Every entry on the Harbor Hub Terminal-Bench 2.1 board spent at least $134.
Which terminal coding agents are open source?
AgenC Core, OpenCode and Pi (MIT), plus Codex CLI, Gemini CLI, Goose and Aider (Apache-2.0). Claude Code and Cursor CLI are closed source.
Which coding agents work with local models?
AgenC Core, Codex CLI (with --oss), OpenCode, Pi, Goose and Aider can use local models through Ollama, LM Studio or an OpenAI-compatible server. Claude Code uses only Claude, and Gemini CLI documents only Gemini.
Which agent scores highest on Terminal-Bench 2.1?
Codex CLI with GPT-6 Astra leads the Harbor Hub board at 87.4%, for $774 a run. Claude Code with Fable 5 follows at 83.8% for $553.
Open source under the MIT license. Source code · All providers · Verify your download