Claude Code and Codex CLI

Use AutoMax from Claude Code and OpenAI Codex, and run AutoMax agents on their logins without an API key.

What you'll learn

How to register AutoMax with Claude Code and Codex so they can drive your test platform, how to run AutoMax's own agent roles through those CLIs instead of an API key, and what each client sees.

Two directions

The CLI drives AutoMax. Register the AutoMax MCP server with the client; then ask it to list projects, run a suite, read a scenario's before/after screenshots, or write a feature proposal.
AutoMax drives the CLI. automax agent … --adapter claude-code|codex runs a role through the CLI's own login. No ANTHROPIC_API_KEY or OPENAI_API_KEY is needed; the CLI's subscription or account pays.

Set up Claude Code

Install and log in

npm i -g @anthropic-ai/claude-code
claude login

Register AutoMax

bun run automax agent install --for claude -p demo-shop -e staging

This writes .claude/agents/automax-<role>.md (one subagent per AutoMax role), CLAUDE.md, AGENT.md and SKILL.md, and runs claude mcp add -s project automax -- npx automax mcp --project demo-shop --env staging (plus the bundled Playwright MCP). Use --no-mcp to skip the registration, or automax mcp install claude --file to write .mcp.json instead of calling the CLI.

Use it

claude
# > use the automax tools to run the demo-shop smoke suite on chromium and summarise failures
# > @automax-healer fix the "Add a product to the cart" scenario from the last run

claude mcp list shows automax and playwright. Every AutoMax tool appears as mcp__automax__<tool>.

Set up Codex CLI

Install and log in

npm i -g @openai/codex
codex login          # ChatGPT account; or set OPENAI_API_KEY

Register AutoMax

bun run automax agent install --for codex -p demo-shop -e staging

This writes AGENTS.md (Codex reads it as project instructions; the five roles become sections), AGENT.md and SKILL.md, and runs codex mcp add automax -- npx automax mcp --project demo-shop --env staging (plus Playwright). Without the CLI on PATH the same entries are merged into ~/.codex/config.toml under [mcp_servers.automax]; other servers in that file are kept.

Use it

codex
# > use the automax tools to list projects and start the api-contract process for demo-shop
codex mcp list

Run AutoMax agents on the CLI logins

bun run automax agent review   -p demo-shop --adapter claude-code
bun run automax agent heal     -p demo-shop --scenario <fingerprint> --adapter codex --max-turns 20
bun run automax agent generate -p demo-shop --goal "checkout with a discount code"   # auto-detects

What the adapters do:

Adapter	Command executed	Tools	Cost reporting
`claude-code`	`claude -p <prompt> --output-format stream-json --mcp-config <temp.json> --strict-mcp-config --allowedTools Read Glob Grep mcp__automax__* mcp__playwright__* --disallowedTools Bash Write Edit … --max-turns <n>`	AutoMax + Playwright MCP, read-only file tools	`total_cost_usd` from the CLI when available
`codex`	`codex exec --json --ephemeral -s read-only -c mcp_servers.automax.command=… -c mcp_servers.automax.args=… <system + prompt>`	AutoMax + Playwright MCP, read-only sandbox	token usage only (Codex reports no dollar cost)

Both adapters keep the AutoMax guarantee: roles can only write proposals (automax proposals list|show|accept). When no adapter is chosen, AutoMax auto-detects: ANTHROPIC_API_KEY → claude CLI → codex CLI → OPENAI_API_KEY → fake, and prints which one it picked.

automax doctor shows both CLIs (installed, logged in) and the full token matrix. Pin a model with AUTOMAX_CLAUDE_CODE_MODEL / AUTOMAX_CODEX_MODEL or --model; by default the CLI's configured model is used.

What each client sees

Client	Reads	Gets from AutoMax
Claude Code	`CLAUDE.md`, `.claude/agents/.md`, `.claude/skills/`, `.mcp.json`	MCP tools, prompts, resources; role subagents; the record/run skills
Codex	`AGENTS.md`, `~/.codex/config.toml`	MCP tools, prompts, resources; role instructions as sections
Cursor, VS Code, Windsurf	`.cursor/mcp.json`, `.vscode/mcp.json`, `.windsurf/mcp.json`	MCP tools, prompts, resources

Next steps

Tokens and keys: what is mandatory, one-of, optional.
Agents: the roles and their proposals.
MCP: tool families, scopes, HTTP transport.

Claude Code and Codex CLI

On this page