AirCode vs Claude Code

At a glance

How the agent workflows differ

Both agents follow the same basic cycle: read the code, plan, edit files, run commands and check the result. The differences are in what each one makes first-class.

AirCode: loops, checks and limits

AirCode is built for work that keeps going while you do something else. aircode loop "make the tests pass" --until "npm test" repeats until the check passes or a limit is reached. The limits are configurable: --max-iter (10 by default), --max-cost (5.00) and --max-minutes (60). With --checkpoint git, AirCode captures progress in Git and can revert agent-made regressions.

aircode watch --check "npm test" runs your check every time you save and starts a fix iteration when it fails. Project-provided MCP servers and check commands stay off until you approve them, and AirCode is designed so commits, pushes, resets and rebases happen only when you ask. /revert undoes the agent’s last edit.

Claude Code: permissions, checkpoints and extensions

Claude Code creates a checkpoint for each prompt, and /rewind (or Esc Esc) restores the code, the conversation or both. Anthropic notes that checkpoints don’t track edits made by Bash commands and aren’t a replacement for version control. Permission modes range from asking before each action to plan mode and accept-edits mode, and an optional sandbox limits file and network access for shell commands.

Its extension points are broad: hooks that run at set points in a session, MCP servers, subagents with their own context windows, skills and plugins, a headless -p mode for scripts, and a GitHub Action that responds to @claude mentions.

Shared ground

Both read a project instructions file (AIRCODE.md for AirCode, which also loads AGENTS.md; CLAUDE.md or AGENTS.md for Claude Code). Both connect MCP servers. Both run headless in CI: aircode run --json or --ndjson, and claude -p with JSON output.

Models and routing

This is the biggest structural difference. Claude Code runs Anthropic’s Claude models and lets you choose among them per session or per role, for example opusplan, which uses Opus in plan mode and Sonnet for execution.

AirCode doesn’t ask you to pick a vendor model for each task. Its requests go through AirLLM, and the default routing model, airllm-auto, matches each step to a suitable model across supported providers. If you prefer a provider family, airllm-anthropic keeps requests within Anthropic’s Claude models and airllm-openai within OpenAI’s. See AirLLM’s routing models and what an AI coding agent router does.

Model choice is also the main cost lever. Anthropic’s own cost guidance suggests Sonnet for most work, Opus for complex reasoning and Haiku for simple subagents. AirCode hands that decision to Air Router instead of to a setting you maintain.

Platforms and availability

Claude Code is available today across terminal, IDE, desktop, web and mobile surfaces. It requires a paid Claude plan or a Console account; the free claude.ai plan doesn’t include it.

AirCode CLI is available today for macOS and Linux, installed with a script or Homebrew, and the Developer Guide also documents a Windows 11 x64 installer. You sign in with aircode login and choose which AirLLM API key it uses. AirCode Studio, a desktop app built on the same agent, is coming soon for macOS, Windows and Linux.

When to choose each

Choose Claude Code if…

• You want Anthropic’s Claude models, managed through an Anthropic plan or Console account.
• Your team wants one agent across the terminal, IDEs, the web, Slack and GitHub Actions.
• You rely on its extension ecosystem: hooks, subagents, skills and plugins.

Choose AirCode CLI if…

• You want each step routed to a suitable model across providers, without choosing one yourself.
• You run long, unattended jobs: loops until the tests pass and fix-on-save, with limits on iterations, time and cost.
• You want one AirLLM key and one bill for the agent and for the other tools you route through Air Router.

You don’t have to choose: Claude Code with Air Router

Claude Code can send its requests to any gateway that speaks the Anthropic Messages API, and Air Router does. Point ANTHROPIC_BASE_URL at https://api.airllm.ai/, put your AirLLM API key in ANTHROPIC_AUTH_TOKEN, and choose a routing model. The Claude Code setup guide lists every variable for macOS, Linux and Windows.

export ANTHROPIC_BASE_URL="https://api.airllm.ai/"
export ANTHROPIC_AUTH_TOKEN="your_airllm_api_key"
# airllm-anthropic stays on Claude models; airllm-auto routes across providers
export ANTHROPIC_MODEL="airllm-anthropic"

Note. Anthropic’s documentation says it doesn’t endorse, maintain or audit third-party gateways, and doesn’t support routing Claude Code to non-Claude models through any gateway. To stay on Claude models, use airllm-anthropic; airllm-auto routes across providers, which is outside Anthropic’s support. Always use your AirLLM key as the token, never a claude.ai login.

Anthropic’s docs also note that with a gateway credential set, subscription usage limits don’t apply and usage is billed per token to whoever owns that credential, which here is your AirLLM account. The Claude Code router guide covers the setup in more depth.

About this comparison. Compiled from each product’s public documentation on October 20, 2026. Details change often, so check the sources below. Product names and trademarks belong to their owners.

Frequently asked questions

Related

Sources

• AirLLM — AirCode CLI
• AirLLM Developer Guide — AirCode CLI and AirCode Studio
• AirLLM Terms — section 10, AirCode software license
• Claude Code docs — Overview
• Claude Code docs — Set up Claude Code
• Claude Code docs — Model configuration
• Claude Code docs — LLM gateway configuration
• Claude Code docs — Checkpointing
• Claude Code docs — Manage costs effectively
• Claude Code license (GitHub)
• Claude pricing