Claude Code 配置检查工具

ccli

See everything your Claude Code installation is configured with.

MCP servers, skills, rules, memory files, and per-project usage,
read straight from disk and shown in one place.

Installation · Quick Start · Commands · Memory Audit · How It Works

Claude Code spreads its configuration across ~/.claude.json, ~/.claude/, .mcp.json, project .claude/ directories, and a managed policy location. Answering "which MCP servers are active here?" or "which CLAUDE.md files actually load in this repository?" means opening several JSON files and knowing the resolution rules.

ccli reads all of it and answers in one command. It works offline and never modifies anything, except for projects clean, which deletes old session data when you ask it to.

Features

• Installation summary: ccli info shows version, auth, model, paths, session and project counts, storage use, and resource counts
• MCP servers: list servers across global and project scopes and inspect their command, URL, and environment, with secrets masked in every output format
• Skills and rules: discover personal, project, and plugin skills, and global and project rules
• Memory audit: see which CLAUDE.md, AGENTS.md, @path imports, and auto memory files load into a session, in order, with their line and byte cost and warnings for files that won't load
• Project usage: last-session cost, tokens, line changes, and per-model cost for every known project
• Session cleanup: delete old session logs and related artifacts, or retire a project entirely, with a --dry-run preview
• Scriptable: every command supports --format json and --format yaml

Installation

Homebrew

brew install jackchuka/tap/ccli

Go

go install github.com/jackchuka/ccli@latest

Binary

Download a prebuilt binary for Linux, macOS, or Windows from the releases page.

Quick Start

Run these from inside a project so that project-scoped config is picked up:

ccli info                                     # installation summary
ccli mcp list                                 # MCP servers across scopes
ccli memory list                              # what loads into a session, in order
ccli projects list                            # per-project cost and sessions
ccli projects clean --older-than 30d --dry-run # preview a cleanup

Commands

Global flags

Cleaning up sessions

# Delete sessions older than 30 days across all projects
ccli projects clean --older-than 30d

# Preview first
ccli projects clean --older-than 30d --dry-run

# Only one project
ccli projects clean my-project --older-than 7d

# Retire a project: every session, its memory, and its config entry
ccli projects clean my-project

--older-than takes a whole number of days (7d, 30d, 90d) and is required unless you name a project. Cleaning removes session logs and the artifacts matched by session UUID: debug logs, telemetry, todos, tasks, file history, and session environment.

Naming a project without --older-than retires it: in addition to every session, it deletes the project's memory/ directory and removes the project from ~/.claude.json. An --older-than sweep never touches memory, because memory belongs to the project rather than to a session.

Memory Audit

ccli memory list               # files loaded at launch
ccli memory list --on-demand   # also files loaded later, when Claude opens a file there
ccli memory get user           # one file, by scope name
ccli memory get CLAUDE.local.md # or by path

ccli memory list shows the managed policy, user, project, and local CLAUDE.md files, every AGENTS.md it finds, their @path imports nested underneath, and the auto memory index, in the order Claude Code concatenates them, with line and byte counts. Under the list it prints totals and the active instruction mode:

  loaded at launch:  5 files · 17L · 315 B
  on demand:         1 file
  instruction files: claude-md-or-agents-md

It warns about:

• a MEMORY.md past its 200-line or 25 KB load cutoff (everything after it is dropped)
• a CLAUDE.md over 200 lines (reduces adherence) or over 4 MiB (skipped entirely)
• broken imports, imports deeper than four hops, cycles, and imports that need external approval
• files excluded by claudeMdExcludes
• auto memory disabled in settings, which means MEMORY.md does not load at all

Each warning appears under its file and again in a summary at the end.

The launch total counts only what actually loads. Excluded, missing, too-deep, cyclic, or mode-skipped files add nothing; a file imported by two parents counts twice because it is expanded twice. The on-demand figure is a count of files found, not a load cost.

The audit follows Claude Code's documented resolution rules rather than instrumenting a running session. /context in a live session remains the ground truth.

AGENTS.md and instruction modes

By default, Claude Code reads AGENTS.md only when there is no project-style CLAUDE.md (CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md) in the working directory or any ancestor, including a bare $HOME/CLAUDE.md. Your ~/.claude/CLAUDE.md and the managed CLAUDE.md don't count. If such a file exists, no AGENTS.md is read, not even in subdirectories. Otherwise a subdirectory's AGENTS.md loads on demand when that directory has no CLAUDE.md of its own. @path imports and claudeMdExcludes apply to AGENTS.md as they do to CLAUDE.md.

ccli lists every AGENTS.md it finds and marks the ones that won't be read, with the reason (for example, the CLAUDE.md that shadows it). Marked files and their imports are left out of the totals.

The mode comes from pluginConfigs["agents-md@builtin"].options.instructionFiles in your user or managed settings; Claude Code ignores it in project and local settings.

An unrecognized value falls back to the default silently, so the totals show claude-md-or-agents-md even if you configured something else. Reading AGENTS.md directly requires Claude Code v2.1.277 or later and is not available on Bedrock, Vertex, or Foundry.

How It Works

ccli reads these files directly. It makes no network calls and only invokes the claude binary for ccli info, to get the version and auth status.

MCP server environment values are masked in text, JSON, and YAML output when the variable name contains KEY, SECRET, TOKEN, PASSWORD, PASSWD, CREDENTIAL, or PRIVATE, or when the value is a URL with a password in it (postgres://user:pass@host/db). Plain URLs such as https://api.example.com are shown as is.

A session counts once per session UUID, whether it exists as a <uuid>.jsonl log, a <uuid>/ directory, or both. Other entries in a project directory, such as memory/, are not counted.

Text output marks each resource's scope (global, project, personal, local, plugin, auto) with a colored bullet.

ccli is built around an agent interface, so support for other coding agents' configuration can be added alongside Claude Code.

License

MIT