Your Claude, Amplified

What Cate adds

One thing: leftenant. It's an MCP server — a sidecar process that gives your Claudes access to project management tools they don't have natively: picking up tasks from the board, creating branches, running quality gates, opening PRs, and handing off between Claudes.

Claude Code spawns leftenant as a child process, just like any other MCP server in your .mcp.json. Cate holds integration credentials in OS secure storage (Keychain on macOS, Credential Manager on Windows) and makes authenticated API calls on leftenant's behalf. Leftenant operates with the same scopes you grant any GitHub, Jira, or Linear integration.

What Cate touches (and what it doesn't)

The pattern: Cate reads everything, writes only to project-level files, and only adds. Never removes or overwrites.

Your login

Cate doesn't read, store, or transmit your Claude Code credentials. OAuth tokens, API keys, enterprise SSO sessions — they flow from your environment directly to the SDK.

Your global settings

Cate tells the SDK to read from the same three settings sources Claude Code uses when you run it from the terminal:

• user — your global ~/.claude/settings.json
• project — the repo's .claude/settings.json
• local — the repo's .claude/settings.local.json

Cate never writes to the global user settings file.

What Cate does write to

Cate writes to the project-level .claude/settings.json — the one inside your repository, not your home directory. Two things happen: leftenant is added to enabledMcpjsonServers, and mcp__leftenant__* is added to the project's permission allow-list so Claudes can call Cate's tools without prompting. Both are additive merges. Your existing project permissions, deny-lists, and other configuration are preserved.

Your enterprise policies

Cate uses the same settings resolution the SDK uses from the terminal and has no mechanism to modify or bypass enterprise policies. Cate passes user in settingSources, never writes to ~/.claude/settings.json, and passes the full host environment through unchanged. The SDK enforces policies at a layer Cate doesn't touch.

Your global skills

Every session launches with the claude_code preset, defined by the Claude Agent SDK itself. It loads the full Claude Code system prompt, including all built-in skills and any global skills you've installed in ~/.claude/commands/. Cate-specific instructions are appended after the preset loads. They don't replace anything.

Your project skills

Cate sets cwd to your project directory, so Claude Code discovers and loads .claude/commands/ exactly as it does from the terminal. Cate doesn't intercept, filter, or modify this discovery.

Your MCP servers

Your MCP servers are read from .mcp.json, environment variables in the config are expanded (so ${FOO_TOKEN} references keep working), and the full server map is handed to the SDK. No servers are filtered out. No configuration is modified.

Cate merges one additional entry into your project's .mcp.json — leftenant. The merge is additive. Your existing servers are preserved.

What to know before you install

Cate adds mcp__leftenant__* to your project's permission allow-list so agents can use its tools without prompting. This is a one-time change that persists for future sessions in that project. And because agents run with your Claude Code permissions in parallel, settings that are fine for interactive single-agent use may warrant a second look for unattended multi-agent runs.

Your Claude. Turned up to eleven.