Most answers to “what are the best AI coding tools?” are rankings, and it’s rarely clear what they rank by. This post doesn’t rank anything. It lines up 18 terminal coding agents under the same headings: who makes it, which command runs it, whether it takes MCP, whether it offers effort levels and how it receives a role. With that you can pick by your own needs.
The details come from the catalog AgentVera uses to launch these tools.
How to choose among the best AI coding tools
Useful questions when weighing a CLI agent:
- Which models can it use? Some tools are tied to one provider; others (like opencode) work with several.
- How do you sign in? With your existing subscription or an API key?
- Does it support MCP? Matters for access to tools like a browser, GitHub or Sentry.
- Does it have effort levels? Some tools offer levels for longer reasoning.
- How does it take lasting instructions? Through a system prompt flag, or files like AGENTS.md in the project?
18 CLI agents at a glance
| Tool | Vendor | Command | MCP (AgentVera) | Effort levels |
|---|---|---|---|---|
| Claude Code | Anthropic | claude | Yes | low, medium, high, xhigh, max |
| Codex | OpenAI | codex | Yes | low, medium, high, xhigh |
| Gemini CLI | gemini | — | — | |
| Copilot | GitHub | copilot | Yes | six levels from minimal to max |
| Cursor | Cursor Agent | cursor-agent | — | — |
| Antigravity | agy | — | low, medium, high | |
| opencode | Any provider | opencode | Yes | The model’s own variants |
| Hermes | Nous Research | hermes | — (its own hermes mcp) | — |
| Qwen Code | Alibaba | qwen | Yes | — |
| Kimi Code | Moonshot AI | kimi | — | low, medium, high, max |
| Kiro | AWS | kiro-cli | — | low, medium, high, xhigh, max |
| Amp | Sourcegraph | amp | Yes | Modes: low, medium, high, ultra |
| Droid | Factory | droid | — | low, medium, high, xhigh, max |
| Goose | Block | goose | Yes | low, medium, high, max |
| Vibe | Mistral | vibe | — | — |
| Crush | Charm | crush | — | — |
| Aider | Open source | aider | — | low, medium, high |
| Cline | Cline CLI | cline | — | low, medium, high, xhigh |
The “MCP (AgentVera)” column shows whether the tool receives AgentVera’s MCP servers and browser on every launch. Some tools may also support MCP through their own settings.
How each tool takes a role
Giving an agent a lasting role (“act like a senior backend engineer, prioritize tests”) helps a lot in parallel work. Tools take it in different ways:
- Appended to the system prompt: Claude Code, Qwen Code (
--append-system-prompt), Goose (--system), Droid (--append-system-prompt). - Codex: as developer instructions.
- Instruction file: an instructions file for opencode; an extra instructions folder (AGENTS.md) for Copilot; an agent file in a new conversation for Kimi.
- Ephemeral instruction: for Hermes, a system instruction passed on every launch and not saved to the session.
- Read-only file: for Aider, a reference file through
--read. - No flag: Cursor, Antigravity, Gemini CLI, Vibe, Crush, Kiro, Amp and Cline offer no extra-instructions flag; use AGENTS.md or rule files in the project for lasting instructions.
Notable details
Claude Code and Codex
The two tools with the broadest support in AgentVera: account chains, context profiles, MCP, the agent browser, conversation copy and /compact. See the Claude Code and Codex pages for details.
opencode
Works with several providers: Anthropic, OpenAI, Google, OpenRouter, local models and the free OpenCode Zen models. You sign in to providers inside opencode itself. The role, MCP servers and the browser are passed to opencode through inline config; conversation copy and /compact are supported.
Gemini CLI and Antigravity
Both come from Google. Gemini CLI uses its own Google sign-in or GEMINI_API_KEY. Antigravity offers model and effort selection.
Copilot
Uses your GitHub sign-in, has the widest effort range (minimal to max) and takes MCP.
Aider and Cline
Aider reads API keys from environment variables or .aider.conf.yml; AgentVera gives each agent its own chat history file. Cline keeps git checkpoints in the project repository.
One tool or several?
Many developers start with one tool and add a second over time: one for planning and building, another for tests or research. For tips on running different tools in one project, see parallel coding agents.
How sign-in differs
The first hurdle in trying a tool is installing it and signing in. Most tools use their own login; AgentVera doesn’t change that, it just launches the CLI:
- Claude Code and Codex: sign in with their own commands (
claude auth login,codex login). AgentVera can manage several accounts with separate config folders. - Cursor: uses its own CLI session; if signed out, the agent’s terminal asks on first launch.
- Antigravity: uses its own Google sign-in.
- opencode: provider sign-in happens in opencode;
/connectin the agent’s terminal. - Hermes: uses the provider and credentials chosen with
hermes model; runhermes setupfirst if needed. - Copilot: GitHub sign-in;
/loginin the terminal if signed out. - Qwen Code: Qwen OAuth or an API key, asked on first launch.
- Goose: uses the provider chosen with
goose configure. - Kimi:
kimi loginor/loginin the terminal. - Kiro:
kiro-cli login. - Vibe: Mistral sign-in or
MISTRAL_API_KEY. - Aider: API keys from environment variables or
.aider.conf.yml.
Model choice and defaults
Most tools use their own default model when none is given. Some use a different concept: in Amp you choose a mode instead of a model, and the mode sets the model, prompt and tools together. Crush’s interactive mode has no model flag; you pick the model inside Crush. In opencode, models are grouped by provider and effort options come from each model’s own list.
Resuming sessions
Picking up where you left off after restarting the app matters on long jobs. For Claude Code and Codex, AgentVera gives every agent a stable key and restarts agents from the CLI’s local history after a normal quit. In Cursor, each agent returns to its own chat; for other tools, resuming depends on what the tool offers.
Checklist
- Do you know which models you can access?
- Are you after a tool that uses your existing subscription, or an API key?
- Do you need MCP servers?
- How will you give a lasting role?
- Installing and signing in to the tool is on you; have you read the provider’s terms?
Wrapping up
The best AI coding tools differ from person to person; the right pick depends on the models you can reach, your subscription and the features you need. The table above puts the basic facts for that decision in one place.
AgentVera runs these 18 tools and a plain terminal in one window. See the full list on the agents page, or start on the free plan.