There's no single best AI coding agent. In practice, the choice comes down to where the agent runs, what account it needs, and whether you can swap the model underneath. Below are short picks by situation, a comparison table of nine agents, and what to check if you bring your own model access. Everything is as of November 14, 2026, with the main sources linked inline.
Quick picks by situation
- You work in the terminal and already pay for Claude or ChatGPT: look at Claude Code (Claude plan) or Codex CLI (ChatGPT plan or API key). Both are made by the model vendors and install with one command: see How to Install Claude Code and How to Install Codex CLI.
- You want an agent inside VS Code or JetBrains: look at Cline or Kilo Code. They're extensions, so you keep the editor you already use. GitHub Copilot also works inside existing IDEs.
- You want the agent and editor from one vendor: Cursor ships its own editor, plus a CLI (
agent) with the same modes. - You want to supply your own model keys, a relay, or a local model: look at Cline, Aider, or OpenCode. Codex CLI also accepts custom providers that implement the Responses API.
- You were on Roo Code or the free tier of Gemini CLI: both changed in 2026. See the notes under the table.
AI coding agents compared
| Agent | Where it runs | License | Account / billing | Can you swap the model? |
|---|---|---|---|---|
| Claude Code | Terminal (plus IDE and desktop apps) | Not open source: the repo's LICENSE.md says "All rights reserved," with use under Anthropic's Commercial Terms | Paid Claude plan, Console, or cloud-vendor credits; not included in free claude.ai | Can point at your own gateway, but Anthropic doesn't support routing it to non-Claude models |
| Codex CLI | Terminal (plus IDE, desktop, and cloud) | Apache-2.0 | ChatGPT sign-in or API key | Custom providers are allowed; the endpoint must implement the Responses API |
| Cursor CLI | Terminal (agent), same modes as the Cursor editor | No public license found | Cursor plan: Free $0, Plus $8, Premium $30, Super $100, Ultra $200 per month (pricing, as of Nov 14, 2026) | Third-party models (Claude, GPT, Gemini) need Premium or above and bill at API list price from prepaid credit; Plus can't use third-party models or your own key |
| GitHub Copilot | IDE extensions plus a CLI | No public license found | Free; Pro $10, Pro+ $39, Max $100 per month; Business $19, Enterprise $39 per seat per month (as of Nov 14, 2026) | Usage allowance depends on the plan |
| Gemini CLI | Terminal | Apache-2.0 | Gemini Code Assist Standard/Enterprise, Code Assist for GitHub, and legacy paid API/Vertex keys | README: free and Google AI Pro/Ultra users moved to Antigravity CLI on June 18, 2026 |
| Cline | VS Code, JetBrains, CLI, desktop, SDK | Apache-2.0 | Your own model keys | README lists Anthropic, OpenAI, Google, OpenRouter, Bedrock, Ollama and more, plus "any OpenAI-compatible API" |
| OpenCode | Terminal, desktop, mobile, SDK | MIT | Your own keys, or ChatGPT Plus/Pro sign-in | Provider docs: model catalog from models.dev; pick a provider with /connect and enter credentials |
| Kilo Code | VS Code, JetBrains, CLI | MIT | README claims 500+ models at the provider's rate, with "zero markup" (vendor's own description) | See README |
| Aider | Terminal | Apache-2.0 | Your own model keys | README: works with almost any LLM, including local models |
How to read the table:
- "Can you swap the model" means changing model providers without changing tools. That decides how tied you are to one vendor.
- Claude and ChatGPT subscription prices weren't captured for this guide, so they aren't listed; check the vendor pages.
- "No public license found" means no public license file turned up, not that none exists.
Changes in 2026. The Roo Code extension was shut down on May 15, 2026; its README points users to ZooCode (a community fork) and Cline. Gemini CLI's free and Google AI Pro/Ultra users were moved to Antigravity CLI (command agy), as shown in the table. If an older list still recommends either for free personal use, it's out of date.
Bringing your own model access
If your models come through an aggregator or your own gateway rather than one vendor's subscription, protocol support is the hard requirement. Three formats matter here:
| Format | Who needs it |
|---|---|
| Anthropic Messages | Claude Code (its gateway requests go to /v1/messages, per the gateway docs) |
| OpenAI Responses | Codex CLI (custom providers must implement it) |
| OpenAI Chat Completions | Tools that accept "any OpenAI-compatible API", such as Cline |
What each tool documents:
- Cline: README lists Anthropic, OpenAI, Google, OpenRouter, Bedrock, Ollama and others, plus "any OpenAI-compatible API."
- Aider: README says almost any LLM, including local models.
- OpenCode: providers come from the models.dev catalog; credentials go in through
/connect; ChatGPT Plus/Pro sign-in also works. - Codex CLI: custom providers, Responses API only (setup walkthrough).
- Claude Code: can point at a gateway; Anthropic doesn't support routing it to non-Claude models (setup walkthrough).
Example: SU8 Codes
SU8 Codes is the product behind this blog. It's an API relay: one key reaches several model providers, and it accepts OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Gemini requests directly (site, as of Nov 14, 2026). Its docs have setup pages for Claude Code, Codex CLI, Cline, Kilo Code, and OpenCode.
SU8 in practice:
- Data path: requests pass through SU8, so there's one more party between you and the model vendor.
- Records: per-request billing details stay online for 31 days; daily summaries and the wallet ledger are kept permanently.
- Team features: no published data on team seats or audit logs.
- Switching: changing providers means changing the client's
base_url. A key can be bound to a channel, and if an upstream misbehaves you can switch channels on the website without touching client config.
FAQ
What is the best AI coding agent?
There isn't one answer. Start from the account you already have, where you work (terminal or editor), and where your models come from, then use the quick picks above.
Is an open-source coding agent safer?
Open source means the code can be audited, not that the defaults are safe. Key storage and network permissions still need to be set per each tool's docs.
How do I avoid having to switch again?
Pick a tool where you supply the model key or endpoint yourself, so the next change is a config edit rather than a new workflow.