Which AI coding agent should you use in the terminal?

Choose a terminal coding agent by its billing model and execution controls before comparing model names. As of October 22, 2026, several familiar recommendations need revision: Gemini CLI has moved individual users to Antigravity CLI, Qwen Code has ended its OAuth free tier, and Cursor has changed ownership and pricing.

Which terminal agents offer the best value?

Claude Code and Codex CLI are sensible first choices if you already pay for Claude or ChatGPT. OpenCode, Aider and Goose suit buyers who want an open-source client and control over inference spending. These are editorial recommendations based on their published features, rather than a performance ranking.

Monthly prices below are in US dollars. A free client does not include free model inference unless stated.

Product	Price	Best for	Limits
Claude Code	Pro $20/month; Max from $100/month	Claude subscribers	Shared five-hour and weekly allowances
OpenAI Codex CLI	Plus $20/month; Pro $200/month; API billing	ChatGPT subscribers	Plus estimates 33 to 168 GPT-5.4 local messages per five hours
Gemini CLI	Client free; paid API or eligible organizational access	Existing Google enterprise users	Individual subscription access discontinued
GitHub Copilot CLI	Free; Pro $10/month; Pro+ $39/month; Max $100/month	GitHub-centered development	Model interactions consume AI Credits
Aider	Client free; separate inference costs	Scripted repository edits	Provider quotas
OpenCode	Client free; optional paid inference	Provider choice	Provider and subscription quotas
Qwen Code	Client free; separate inference costs	Qwen and compatible APIs	OAuth free tier discontinued
Amp	Hobby free; Individual $20/month	Subscription reuse and remote agents	Hosted compute and inference billed separately
Goose	Client free; separate inference costs	MCP-based automation	Provider quotas and local compute
Cursor CLI	Free; Plus $8/month; Premium $30/month	Cursor and Grok users	Weekly shared pool; third-party models extra
Warp Agent CLI	$0/month; metered inference or BYOK	Terminal and cloud handoff	No fixed free inference allocation promised

Treat those ceilings differently. Claude pools usage across its apps and Code. Codex describes task-dependent ranges. GitHub meters credits, while provider-neutral clients pass model costs to your chosen account.

Claude Code and Codex: start with your existing subscription

Claude Pro includes Claude Code for $20 monthly, or $200 billed annually. Max starts at $100 monthly. Anthropic applies rolling five-hour limits plus weekly limits, with no fixed message count because model choice and task complexity change consumption.

Claude Code is proprietary and centers on Claude models, including the Opus, Sonnet and Haiku families. Authentication options include Claude subscriptions and supported commercial API platforms. Install through Anthropic's native installer, or use npm install -g @anthropic-ai/claude-code.

Its permission system controls tool actions; Bash sandboxing adds restrictions beyond approval prompts. MCP connects external tools. For scripts, claude -p runs without the interactive interface. Configure permissions before a CI run rather than relying on someone to approve a terminal prompt.

Codex CLI is Apache-2.0 licensed. Install with npm install -g @openai/codex. OpenAI includes it with ChatGPT Plus and Pro, and offers token-based API-key access. Free and Go access is described as a limited-time offer, so avoid making it the basis of a long-term budget.

OpenAI's Codex pricing page lists GPT-5.4, GPT-5.3-Codex and GPT-5.4-mini. For GPT-5.4, Plus estimates 33 to 168 local messages per five hours; additional weekly limits may apply. OpenAI says GPT-5.4-mini permits up to 3.3 times higher local-message usage. Codex supports MCP and codex exec for automation, with explicit sandbox and approval settings.

Gemini CLI and Qwen Code: old free-tier advice has expired

Gemini CLI stopped serving individual free, Google AI Pro and Ultra accounts on June 18, 2026. Google directs those users to Antigravity CLI. Eligible Gemini Code Assist organizational licenses and paid API access remain supported. The Apache-2.0 project continues, but its README warns that it is not optimized for Gemini 4 Argon.

That change makes the old consumer recommendation of 1,000 free Gemini CLI requests per day unsuitable for a new purchase decision. Existing enterprise users should check their organization's entitlement; new individual users should evaluate Antigravity under its own terms.

Gemini CLI installs with npm install -g @google/gemini-cli. It supports MCP, headless prompts through gemini -p, approval modes and optional sandboxing through gemini -s. Sandboxing is off by default.

Qwen Code's OAuth free tier ended on April 15, 2026. Its current authentication choices include QwenCloud plans, Alibaba ModelStudio and compatible provider APIs. The Apache-2.0 client also supports OpenAI, Anthropic and Gemini protocols, plus local endpoints such as Ollama or vLLM.

Install with npm install -g @qwen-code/qwen-code. Qwen supports MCP and qwen -p automation. Its approval modes range from read-only planning to YOLO. An important trade-off: the documented sandbox mode disables MCP and several other integrations. Test the configuration you intend to deploy, rather than assuming every feature works together.

Copilot, Cursor and Warp: check what the account includes

GitHub Copilot CLI is the standalone copilot application, replacing the retired gh copilot extension. Install it with npm install -g @github/copilot. It uses a proprietary license, despite having a public repository.

The current plans include CLI access on Free, Pro, Pro+ and Max. Pro costs $10 monthly, Pro+ $39 and Max $100. GitHub's AI Credits are worth $0.01 each; a plan's dollar credit allowance is distinct from its subscription price. Free's 50 chat requests should not be read as a guaranteed allocation of 50 autonomous coding tasks.

Model access varies by plan and client. Use /model to inspect your choices. Copilot supports MCP, configurable tool permissions, sandboxing and programmatic operation, with token authentication for CI.

Cursor became part of SpaceX on August 14, 2026. New subscriptions differ from legacy Pro plans: Plus costs $8 monthly and excludes third-party models; Premium costs $30 and permits those models at separate API list prices. Included Cursor-model usage refreshes weekly and is shared with Grok. Grok and Composer models draw from that pool.

Cursor CLI is proprietary. Install through Cursor's installer and run agent. It shares MCP configuration with the editor, supports sandbox settings, and offers agent -p with structured output for scripts.

Warp's client became open source in April 2026, with AGPLv3 covering most repository code and MIT covering its UI framework. Distinguish that client license from hosted services and the separately distributed Agent CLI. The current command is warp; legacy Oz support was scheduled only through September 2026.

Install the Agent CLI through Warp's installer or brew install --cask warp-agent-cli. It supports provider selection, BYOK and MCP. Permission cards govern commands and edits. API-key authentication supports browserless machines; its reference centers on interactive sessions and cloud handoff rather than a documented one-shot print mode.

OpenCode, Aider and Goose: choose your own inference

OpenCode is MIT licensed and supports hosted providers and local models. Its free software and selected free models are separate from paid provider accounts or optional OpenCode inference services.

Watch the version boundary. OpenCode 1 uses npm install -g opencode-ai; OpenCode 2 uses npm install -g @opencode/cli@latest. Both claim the opencode command. Match installation, permission configuration and automation instructions to the same generation.

OpenCode supports local and remote MCP servers. Version 2 permissions use allow, deny and ask rules. These are action policies; use an isolated execution environment when you need a stronger boundary. Version 1 documents opencode run and a headless HTTP server for automation.

Aider is Apache-2.0 licensed and works with Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter and local providers through its model integrations. Install with python -m pip install aider-install, then run aider-install. There is no bundled inference allowance.

For a scripted change, aider --message runs a task and exits. Confirmation controls and dry-run operation are useful, but they should not be mistaken for OS isolation. Aider's official feature documentation does not offer a first-party MCP integration. Prefer another client if MCP is a requirement.

Goose is Apache-2.0 licensed under its current repository license, with the project now housed at aaif-goose/goose. It accepts multiple hosted providers and local Ollama models. Install using the project's CLI installer or brew install --cask block-goose.

Goose builds its extension system around MCP. It can run local commands and modify files, so configure its permission controls before giving it an unfamiliar repository. For unattended work, goose run -t accepts a task, with structured output options for automation.

Amp separates agent access from compute costs

Amp's Hobby tier is free when you bring your own compute and model subscriptions or keys. Its September 13, 2026 announcement removed the monthly-plan requirement for using a ChatGPT subscription and removed ordinary BYOK token fees and limits outside Enterprise.

Individual costs $20 monthly and includes 45,000 minutes of hosted-orb time. Orbs are remote computers; inference is a separate consideration. Self-hosted runners are free, while Amp-hosted inference can be purchased at usage-based prices.

Amp is proprietary. It supports Claude and GPT subscriptions or keys, with model routing and pinning. Install using Amp's CLI installer. MCP is supported, and amp -x runs one turn and exits for scripts.

Pay attention to its default execution policy: Amp's tools documentation says it does not ask for approval by default. For an untrusted repository, use an isolated environment or an explicit policy plugin. Workspace MCP servers require approval, but that is separate from approving each shell command.

Choose by budget, model preference and operating requirements

For a $20 monthly budget, start with the subscription you already use. Pick Claude Code for Claude access or Codex for ChatGPT inclusion. Avoid paying for both until one tool's limits or workflow creates a specific need.

For a free client with paid API flexibility, choose OpenCode for provider switching, Aider for bounded scripted edits, or Goose for MCP-heavy automation. Local inference removes a hosted token bill, but still needs suitable hardware and a model capable of your task.

For a strict open-source requirement, shortlist Codex, OpenCode, Aider, Goose, Gemini CLI and Qwen Code. Check model terms separately: an open-source agent does not make its hosted model open source. Warp's open client also deserves consideration, with its AGPL obligations reviewed against your intended use.

For GitHub-centered teams, evaluate Copilot CLI against actual credit consumption. Cursor users should inspect weekly shared usage and third-party charges before adding CLI workloads. Choose Warp when interactive terminal continuity and cloud handoff matter more than a one-shot scripting interface.

For CI, prefer a documented non-interactive entry point, explicit spending limits and disposable execution environments. Keep approval bypasses out of trusted defaults, restrict credentials, and review the generated diff before merging.