Claude Code × PingAI Workflows: Setup, Memory, Skills, MCP, and CI/CD

Claude Code is Anthropic's official terminal AI coding agent (CLI). It reads and modifies local codebases, runs commands, and performs Git operations. Connected through the PingAI gateway, a pk- key is all you need — no official Claude subscription required.

Native install (recommended, auto-updates in the background):

Windows PowerShell:

npm (requires Node.js 22+):

WinGet (Windows):

Verify:

On native Windows, install Git for Windows so Claude Code can use the Bash tool.

Connecting to PingAI requires two variables: the base URL and the API key.

Option 1 (settings file, applies globally): write to ~/.claude/settings.json (Windows: %USERPROFILE%\.claude\settings.json):

Option 2 (shell export, current terminal only):

PowerShell:

Note: do not put credentials in a project .claude/settings.json (it would be committed and shared with the repo); the settings file takes precedence over shell exports. You can also use npx pingai wizard to write it automatically.

Verify the connection after configuring:

1. 1Run claude in your project directory
2. 2Type /status — an Anthropic base URL line pointing at https://api.pingai.hk, plus an Auth token / API key line, means the gateway config is active
3. 3Send any prompt — a normal reply confirms the connection

To verify with curl directly (without opening Claude Code):

A reply starting with {"id":"msg_ and containing "content":[...] means success; 401 means the key was rejected.

Every Claude Code session starts with a fresh context window; cross-session knowledge lives in CLAUDE.md (instructions you write). Put project instructions in ./CLAUDE.md or ./.claude/CLAUDE.md so they version-control with the team; put personal preferences in ~/.claude/CLAUDE.md. Keep each file under 200 lines and use verifiable, concrete phrasing like "Use 2-space indentation" or "Run npm test before committing". Use /init to generate an initial CLAUDE.md, /context to confirm what is loaded, and /memory to edit.

In a PingAI project you can even write model preferences and cost rules into CLAUDE.md so Claude Code follows them on every run.

For large projects, split instructions by topic into .claude/rules/ (e.g. testing.md, api-design.md, security.md); all .md files are discovered recursively. Rules with paths frontmatter load only when Claude reads a matching file, saving context:

auto memory is on by default: Claude records build instructions, debugging lessons, and code style preferences to ~/.claude/projects/<project>/memory/, where MEMORY.md is the index. All worktrees of the same git repo share one memory directory, stored locally only.

A Skill is a directory containing a required SKILL.md (YAML frontmatter + instruction body); the directory name is the command, invocable as /skill-name. Locations: personal ~/.claude/skills/<name>/SKILL.md (across projects), project .claude/skills/ (shared, committed). Minimal example:

Argument placeholders: $ARGUMENTS (all arguments), $ARGUMENTS[N] / $N (by position). Dynamic context injection !command`` runs a command before sending to Claude and replaces it with the output. Unlike CLAUDE.md, skill bodies load only when invoked — a good fit for long procedure docs.

A Subagent is a specialist with its own context window, system prompt, and permissions; it returns only a summary to the main conversation, protecting the main context. Three built-in types: Explore (read-only, good at searching / analyzing codebases), Plan (read-only research in plan mode), and general-purpose (a complex multi-step assistant with full tools). Custom subagents live in .claude/agents/:

Cost control: model: haiku routes simple tasks to a cheaper model; permissionMode: acceptEdits lets the subagent accept file edits automatically.

MCP (Model Context Protocol) lets Claude Code talk directly to external tools / databases / APIs. Add a server:

Manage: claude mcp list, claude mcp get <name>, claude mcp remove <name>; the in-session /mcp panel shows connection status. Three scopes: --scope local (default, current project only), --scope project (writes .mcp.json to share with the team), --scope user (across all projects). Security: only connect servers you trust; MCP tools go through the standard permission system and can be restricted with permissions.deny.

Hooks are user-defined shell commands that Claude Code runs deterministically at fixed lifecycle points (not driven by LLM decisions), used to enforce rules and automate workflows. Core events: SessionStart, UserPromptSubmit, PreToolUse (can block before a tool runs), PostToolUse (after success), PermissionRequest, Stop, SessionEnd. Configure the hooks block in settings:

In practice: auto-run prettier formatting on PostToolUse, or protect .env on PreToolUse (exit 2 blocks it). Scope depends on location: ~/.claude/settings.json (global), .claude/settings.json (shared, committed), .claude/settings.local.json (personal).

Non-interactive mode takes a prompt with -p, ideal for scripts and CI:

For structured output use --output-format json (includes result / session_id / per-model cost total_cost_usd). In CI, add --bare to skip hooks / skills / plugins / MCP / CLAUDE.md auto-discovery so every machine behaves identically.

On GitHub Actions use the official anthropics/claude-code-action@v1, add PINGAI_API_KEY (a pk- key) to Secrets, and point at PingAI via environment variables:

Cost control: a sane --max-turns prevents runaway iterations, add a workflow-level timeout, and use GitHub concurrency limits.

Combining everything above, here is a complete example — every PR is automatically code-reviewed by Claude Code, with the cost billed to your PingAI credits quota:

1. 1Create .github/workflows/pr-review.yml in the GitHub repo
2. 2Add the Secret PINGAI_API_KEY (a pk- key)
3. 3The workflow: PR opens → run claude-code-action pointed at PingAI → post a comment on the PR

Key points: CLAUDE.md is the coding-standard source Claude follows in CI; --model claude-sonnet-5 pins the main model; --max-turns stops runaway loops from burning credits. When the daily run finishes, the cost is fully visible on the Console usage page.