AI Skill Hub 推荐使用:Codex/Claude Code skill + CLI + MCP server 是一款优质的MCP工具。AI 综合评分 7.5 分,在同类工具中表现稳健。如果你正在寻找可靠的MCP工具解决方案,这是一个值得深入了解的选择。
Codex/Claude Code skill + CLI + MCP server 是一款遵循 MCP(Model Context Protocol)标准协议的 AI 工具扩展。通过 MCP 协议,它可以让 Claude、Cursor 等主流 AI 客户端直接访问和操作外部工具、数据源和服务,实现 AI 能力的无缝扩展。无论是文件操作、数据库查询还是 API 调用,都可以通过自然语言在 AI 对话中直接触发,极大提升生产效率。
Codex/Claude Code skill + CLI + MCP server 是一款遵循 MCP(Model Context Protocol)标准协议的 AI 工具扩展。通过 MCP 协议,它可以让 Claude、Cursor 等主流 AI 客户端直接访问和操作外部工具、数据源和服务,实现 AI 能力的无缝扩展。无论是文件操作、数据库查询还是 API 调用,都可以通过自然语言在 AI 对话中直接触发,极大提升生产效率。
# 方式一:通过 Claude Code CLI 一键安装
claude skill install https://github.com/1939869736luosi/codex-sessions-manager
# 方式二:手动配置 claude_desktop_config.json
{
"mcpServers": {
"codex-claude-code-skill---cli---mcp-server": {
"command": "npx",
"args": ["-y", "codex-sessions-manager"]
}
}
}
# 配置文件位置
# macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
# Windows: %APPDATA%/Claude/claude_desktop_config.json
# 安装后在 Claude 对话中直接使用 # 示例: 用户: 请帮我用 Codex/Claude Code skill + CLI + MCP server 执行以下任务... Claude: [自动调用 Codex/Claude Code skill + CLI + MCP server MCP 工具处理请求] # 查看可用工具列表 # 在 Claude 中输入:"列出所有可用的 MCP 工具"
// claude_desktop_config.json 配置示例
{
"mcpServers": {
"codex_claude_code_skill___cli___mcp_server": {
"command": "npx",
"args": ["-y", "codex-sessions-manager"],
"env": {
// "API_KEY": "your-api-key-here"
}
}
}
}
// 保存后重启 Claude Desktop 生效
Codex Desktop now includes a delete action for archived chats. Local testing shows that it removes the main session file and some thread rows, but may still leave session index rows, execution logs, and desktop state references behind.
codex-sessions-manager is a local Codex session audit and cleanup tool. It works as a Skill (Claude Code / Codex), a CLI, and an MCP server — all sharing the same core. Use it to inspect what remains under ~/.codex, audit leftovers after the official UI delete/archive flow, clean hidden local residues by exact session ID, and verify that deletion actually left no local orphans.
| Feature | What it does |
|---|---|
| **List & filter** | By project, status, time range, source metadata, model provider, and model; group by project |
| **Source summary** | Read-only sourceKind summary while preserving raw source, thread_source, model_provider, model, and agent_role |
| **Split title sources** | Lists show the Codex UI-searchable title by default; detail output shows session_index, SQLite, and first-message title differences |
| **Export** | Backup any session to JSON before you touch it |
| **Delete** | Permanent or recoverable trash — your choice |
| **Residue audit** | Read-only report for raw rollout files, shell snapshots, session indexes, history, SQLite rows, global-state refs, thread edges, family status, and broken parent/child links |
| **Root residue scan** | Read-only root-level scan for likely leftover IDs, without requiring a session ID first |
| **Root delete preview** | Read-only batch delete preview for root residue candidates, without requiring you to list session IDs by hand |
| **Trash & Restore** | Full snapshot saved; restore checks for SQLite key conflicts before writing |
| **Verify** | Reports any remaining files, index rows, or DB records |
| **Cleanup** | Remove stale index entries without touching raw data |
| **Health check** | doctor command for full root diagnostics |
| **MCP server** | AI agents (Claude Code, Codex, Kiro) manage sessions directly |
| **Session family** | Read-only parent, child, ancestor, descendant, sibling, subagent, /fork, /side, and impact views; human output uses short source labels unless --full is used |
| **Side conversations** | Parent and child sessions stay separate; delete/export/verify never recurses automatically |
npm install -g codex-sessions-manager
```bash
codex-sessions list [--status active|archived] [--limit N] [--project TEXT]
codex-sessions list --updated-after 2026-04-01 --updated-before 2026-04-30
codex-sessions list --group-by project
codex-sessions list --source-kind cli --model-provider openai
codex-sessions list --source mcp --thread-source mcp
codex-sessions list --agent-role subagent --agent-nickname helper
codex-sessions sources [--json]
codex-sessions projects
codex-sessions doctor [--json]
codex-sessions show <session-id>
codex-sessions family <session-id> [--json] [--children|--parents|--subagents|--impact] [--full] [--source-kind KIND]
codex-sessions audit <session-id> [--json]
codex-sessions audit-root [--json] [--limit 50] [--status STATUS...] [--source SOURCE...] [--all]
codex-sessions preview-root [--json] [--limit 50] [--status STATUS...] [--source SOURCE...] [--all]
codex-sessions export <session-id> [--output ./backup.json]
codex-sessions delete <session-id...> [--trash] [--yes]
codex-sessions trash-list
codex-sessions restore <trash-id-or-session-id> --yes
codex-sessions purge <trash-id-or-session-id> --yes
codex-sessions cleanup-stale [--yes]
codex-sessions cleanup-index <session-id...> [--yes]
codex-sessions verify <session-id...> [--json]
Safety first: All destructive commands require --yes. Without it, you only get a preview. Run a separate preview for the exact session IDs first; family, impact, audit-root, and preview-root never count as permission to delete.
export and trash bundles are recovery data, not previews. They may include full global-state exact-key values such as prompt-history content. Human delete previews show only path, rule, shape, and byte counts.
Duplicate trash entries: restore does not delete the trash entry. If a restored session is moved to trash again, trash-list can show multiple recoverable copies for the same session ID. This is normal trash state, not live residue. A newer copy does not replace an older one. When one session ID has multiple trash entries, confirmed restore and purge refuse the session ID and require the exact trashId. Do not auto-purge duplicates. purge permanently removes only the selected trash entry and never touches the live session.
Use audit after the official Codex UI delete/archive flow when you need a clear local residue report. It is read-only. It reports whether the raw rollout file, shell snapshot, session_index, history, SQLite records, known global-state refs, P11 exact-key global-state refs, unknown global-state refs, and thread_spawn_edges are still present. It also reports family membership and broken parent/child links. If anything remains, the suggested next command is a preview-only delete command; nothing is deleted unless you add --yes.
Use audit-root when you do not already have the session ID. It scans the whole Codex root and lists likely residue candidates by risk: broken parent/child edges, missing rollout files with unknown global-state refs, SQLite-only rows, shell snapshots, index-only rows, and other partial leftovers. It is read-only, defaults to --limit 50, does not print transcript content, and recommends a per-session audit command for each candidate. Add --all only when you intentionally want complete non-residue sessions included too.
audit-root supports display-only filters. Matching candidates are not a deletion list or a deletion recommendation; audit them one by one or inspect a read-only preview before choosing any cleanup:
--status risky-global-state--status global-state-exact-key--status db-only--status broken-family--status partial-residue--status global-state-unknown--source global-state-unknown--source global-state-exact-key--source global-state-known--source sqlite--source session-index--source history--source shell-snapshot--source thread-spawn-edgesYou can pass --status or --source more than once. Multiple values of the same kind use OR. Combining status and source uses AND. These filters only narrow what is shown. A matching candidate still needs per-session audit or delete preview before any cleanup decision, and it does not mean the candidate should be deleted.
Human and JSON output include a summary: filters, totalCandidatesBeforeFilter, totalCandidatesAfterFilter, returnedCandidates, limit, byStatus, and bySource. The byStatus and bySource counts are computed after status/source filters and before limit.
Use sources when you need a read-only overview of where sessions came from. It groups by inferred sourceKind, raw source, thread_source, model_provider, model, and agent_role. sourceKind can be subagent, mcp, vscode, cli, exec, or unknown. The raw source value is still kept in JSON output and shown in human output, because sourceKind is only this tool's inferred category.
list supports the same source-facing filters: --source-kind, --source, --thread-source, --agent-role, --agent-nickname, --model-provider, and --model. Filters combine with AND across different fields. Repeating the same field uses OR. MCP list_sessions accepts the same fields, and MCP summarize_sources returns the same summary shape as CLI sources --json.
Important source limits:
source=vscode is a raw Codex thread source label. It should not be treated as proof that the chat came from the VS Code IDE.cli, mcp, vscode, or exec are unknown, not automatically Desktop.source=mcp means the thread was recorded with that source. It is not a log of every MCP tool call inside the conversation.model_provider is only displayed and filtered here. This tool does not repair provider identity or rewrite provider history.Use preview-root when you want a read-only batch delete preview for the same candidates audit-root would select. It reuses the same status/source filters and conservative default --limit 50, then summarizes what a read-only preview would touch across rollout files, shell snapshots, session_index, history, SQLite, known global-state refs, P11 exact-key global-state refs, unknown global-state refs, and thread_spawn_edges. It does not delete, does not rewrite JSONL, SQLite, shell snapshots, or global-state, does not accept --yes, does not recommend deleting any session, and does not recursively add parent, child, or family sessions. A preview-root result is not a deletion recommendation; it only shows what would be touched if you later choose explicit delete commands. Actual deletion still requires a separate explicit-ID delete preview and then a separate delete ... --yes command.
该项目是一个开源的MCP工具,支持Codex/Claude Code skill、CLI和MCP server,用于本地会话管理。虽然项目质量较高,但仍然需要进一步的测试和验证。
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ Apache 2.0 — 宽松开源协议,可商用,需保留版权声明和 NOTICE 文件,含专利授权条款。
总体来看,Codex/Claude Code skill + CLI + MCP server 是一款质量良好的MCP工具,在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态,建议收藏备用,结合自身场景选择合适时机引入使用。
| 原始名称 | codex-sessions-manager |
| 原始描述 | 开源MCP工具:Codex/Claude Code skill + CLI + MCP server for local session management — 4-laye。⭐9 · TypeScript |
| Topics | mcpai-agentclaude-codecleanupclicodextypescript |
| GitHub | https://github.com/1939869736luosi/codex-sessions-manager |
| License | Apache-2.0 |
| 语言 | TypeScript |
收录时间:2026-05-23 · 更新时间:2026-05-30 · License:Apache-2.0 · AI Skill Hub 不对第三方内容的准确性作法律背书。
选择 Agent 类型,复制安装指令后粘贴到对应客户端