Skip to main content

Use Claude Code, Codex CLI, or Gemini CLI as Atmos AI Providers

· 5 min read
Andriy Knysh
Andriy Knysh
Principal Architect @ Cloud Posse

Atmos AI now supports CLI providers — invoke your locally installed Claude Code, OpenAI Codex, or Gemini CLI as AI backends. No API keys needed. Just use your existing subscription.

Why This Matters​

Until now, using Atmos AI required purchasing API tokens from a provider and configuring keys. Many developers already have Claude Code or Codex installed with active subscriptions (Claude Max, ChatGPT Pro, or Gemini's free tier). CLI providers let you reuse that investment:

  • No API keys — the CLI tool handles auth via its own subscription
  • No per-token billing — included in your existing plan
  • Full MCP support — Claude Code and Codex CLI can use all your configured MCP servers

Quick Start​

atmos.yaml
ai:
  enabled: true
  default_provider: "claude-code"  # or "codex-cli" or "gemini-cli"
  providers:
    claude-code:
      max_turns: 10
brew install --cask claude-code && claude auth login
atmos ai ask "What did we spend on EC2 last month?"

That's it. Atmos detects the binary, generates an MCP config with auth wrapping, and passes everything to Claude Code. You get answers from real AWS data using your subscription.

Available CLI Providers​

ProviderBinarySubscriptionMCP Support
Claude CodeclaudeClaude Pro/Max ($20-200/mo)Full
OpenAI CodexcodexChatGPT Plus/Pro ($20-200/mo)Full
Gemini CLIgeminiGoogle account (free tier)Blocked for personal accounts

All three providers work for prompt-only queries. For MCP-enabled workflows (AWS billing, security, IAM, etc.), use Claude Code or Codex CLI.

MCP Pass-Through​

When mcp.servers is configured in atmos.yaml, Atmos automatically passes MCP servers to the CLI tool. Each provider uses its native config format:

  • Claude Code — temp .mcp.json via --mcp-config
  • Codex CLI — ~/.codex/config.toml (backup/restore after exit)
  • Gemini CLI — .gemini/settings.json in the current working directory

Auth-requiring servers are wrapped with atmos auth exec -i <identity>. Toolchain PATH and ATMOS_* env vars are injected automatically.

See It in Action​

Outputs below are from a production setup with 8 AWS MCP servers configured.

Claude Code — Security Posture​

$ atmos ai ask "What is our security posture in us-east-2 region?"

ℹ MCP servers configured: 8 (config: /tmp/atmos-mcp-config.json)
ℹ AI provider: claude-code
👽 Thinking...

  ## Security Posture Summary for us-east-2

  ### ✅ Enabled Security Services (2/6)

   GuardDuty - Fully operational

    • Status: ENABLED
    • Finding frequency: Every 6 hours
    • Active data sources: CloudTrail, DNS Logs, VPC Flow Logs,
      S3 Logs, EKS Audit Logs, RDS Login Events
    • Current findings: None (clean)

   Inspector - Partially enabled

    • Status: ENABLED but no scan types active

  ### ❌ Disabled Security Services (4/6)

    1. IAM Access Analyzer - Not configured
    2. Security Hub - Not enabled
    3. Trusted Advisor - Error checking status
    4. Macie - Not enabled

  ### Key Recommendations

   High Priority:
    1. Enable Security Hub for centralized findings aggregation
    2. Enable IAM Access Analyzer
    3. Enable at least one Inspector scan type

   Your security posture is moderate — you have basic threat detection
   enabled but are missing several important security services.

Codex CLI — EC2 Billing​

$ atmos ai ask "What did we spend on EC2 last month in each region?"

ℹ MCP servers configured: 8 (in ~/.codex/config.toml)
ℹ AI provider: codex-cli
👽 Thinking...

   For last month, I used the previous calendar month:
   March 1, 2026 through April 1, 2026.

   EC2 spend by region:

    • us-east-2: $88.10

   Cost Explorer returned only us-east-2 for Amazon Elastic Compute
   Cloud - Compute, so that appears to be the only region with EC2
   spend in that period. AWS also marked the result as Estimated,
   which is normal on April 1, 2026 while billing finalizes.

Both providers automatically selected the right MCP server (aws-security and aws-billing) and returned answers from real AWS data — no manual server selection needed.

Configuration​

You can mix CLI and API providers in the same config and switch between them:

atmos.yaml
ai:
  default_provider: "claude-code"
  providers:
    claude-code:
      max_turns: 10
    codex-cli:
      full_auto: true
    anthropic:
      model: "claude-sonnet-4-6"
      api_key: !env "ANTHROPIC_API_KEY"
# Uses claude-code (default)
atmos ai ask "What stacks do we have?"

# Override to codex-cli
atmos ai ask --provider codex-cli "What did we spend on EC2?"

# Override to API provider
atmos ai ask --provider anthropic "Describe the vpc component"

Try It​

Example: AI with Claude Code

Complete example with Claude Code CLI provider, AWS MCP servers, and automatic auth.

Try it out​

Copy the example into a new directory and make it your own:

atmos init examples/ai-claude-code
cd ai-claude-code

Install and authenticate Claude Code as described in the README before making a request.

atmos ai ask "What stacks do we have?"
Explore the example →Initialization and repository settings

Learn More​