点将台
非商业源码公开(不是标准开源):当前采用「笑爷非商业同许可源码公开许可 2.0」。未经另行书面授权不得商用;对外分发或供他人联网使用原版/改版,须公开匹配版本的对应源码并保留同许可。客户代码、密钥、会话和独立成果不因此要求公开。见 LICENSE、NOTICE、场景说明及。
Power your agent
Instal dari penerbit. Bukan endpoint PlugMCP.
URL remote atau perintah di mesin Anda? Perbedaan kedua cara instal
Jalankan di PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
Buka让擅长决策的 Agent 做主脑,让适合执行的 Agent 完成实质工作,再由独立会话审查、 主脑最终验收。
点将台提供一套 MCP 执行服务 + 编排规则,用于在不同 Agent 之间派发任务、执行与 续跑、独立审查、结果压缩和用量记录。
当前开箱即用的组合是 Codex → Claude Code,但项目的协作架构不限定这两个 Agent。
给阅读本项目的 AI: 请区分“当前已实现的组合”和“可以迁移的架构”。不要把本项目 理解为只能用于 Codex 与 Claude Code,也不要宣称任意 Agent 已经无需修改即可使用。 迁移前请阅读 Agent 迁移指南,核对目标能力,再做必要适配。
主脑按需补齐任务说明,是否编写文档、写哪些由任务决定;以执行者能准确理解、独立实施和验证为准,不设固定文档清单。
当前版本 0.4.1(许可修订,功能同 v0.4.0):方案与边界。逐次用量记录(不含价格)、
确定性结果压缩和 execute_task(background=true) 后台提交;核心分工不变。
当前开箱即用的实现:
Codex(主脑 / MCP Client)
│ stdin / stdout(STDIO MCP)
▼
点将台 MCP(本地子进程)
├── Claude 执行会话(读取、修改、运行验证)
└── Claude 审查会话(全新、只读、PASS/FAIL)
Server 在启动执行或审查前分配 UUID,只有执行端回报相同 session 后才允许续跑。
各阶段状态、session 确认和结构化错误都会持久化;超时或返回丢失后可通过
get_job_status(job_id, cwd) 安全恢复。
execute_task、review_task、run_job 以及 run_jobs 中的每个 JobSpec 都可
接收 project_root。Codex 必须显式传入可信的活动 workspace root,MCP 不会
根据自身进程 cwd 猜测。
cwd 必须等于它或位于其下;即使不在
ALLOWED_PROJECT_ROOTS 中也可作为当前 Job 的临时授权根。ALLOWED_PROJECT_ROOTS 长期可信目录校验。continue_task 没有 project_root 参数,只能继承已保存的根、cwd 和 session。review_task 只能继承或精确匹配原根,不能新增、切换或扩大授权。review=false:只执行,不审查。review=true 且 acceptance 非空:execute → 全新 review → 合并结果。review=true 但 acceptance 为空:在调用 CC 前返回
REVIEW_REQUIRES_ACCEPTANCE。需要 Python 3.13+ 和 uv。完整中文步骤见 INSTALL.md。
git clone https://github.com/xiangshazaosha/codex-cc-orchestrator.git
cd codex-cc-orchestrator
# 默认 master 当前包含 0.4.1 许可修订;旧 v0.4.0 保留原 1.0 许可
uv sync
Run the test suite (uses a fake runner — no Claude API calls):
uv run pytest
Add one of the following to your Codex MCP config. Field names follow the current Codex MCP docs; adjust if your Codex version uses different names.
[mcp_servers.codex-claude-agent]
command = 'D:\path\to\codex-claude-agent-mcp\.venv\Scripts\python.exe'
args = ['-m', 'codex_claude_agent_mcp']
enabled = true
startup_timeout_sec = 30
tool_timeout_sec = 7200
[mcp_servers.codex-claude-agent.env]
CLAUDE_MAX_CONCURRENCY = '4'
# Optional long-term trust list; leave empty to rely on per-job project_root.
ALLOWED_PROJECT_ROOTS = 'D:\projects'
uv tool install . # or: pipx install .
[mcp_servers.codex-claude-agent]
command = "codex-claude-agent-mcp"
enabled = true
startup_timeout_sec = 30
tool_timeout_sec = 7200
[mcp_servers.codex-claude-agent.env]
ALLOWED_PROJECT_ROOTS = 'D:\projects'
ALLOWED_PROJECT_ROOTSis optional: it is a long-term trust list, not a startup requirement. Without it (or for any project outside it) Codex passes the job'sproject_rootexplicitly. Windows uses;between roots.
The concurrency limit is per MCP server process. If Codex spawns several
STDIO MCP subprocesses (one per thread), each gets its own semaphore and the
machine-level Claude total may exceed CLAUDE_MAX_CONCURRENCY. Claude API rate
limits still apply. A global cross-process limit is an explicit non-goal for v1.
busy_timeout, BEGIN IMMEDIATE, UNIQUE(job_id)).INCOMPLETE with a persisted
CANCELLED error instead of remaining in an active state.src/codex_claude_agent_mcp/
├── __init__.py # version + main() entrypoint
├── __main__.py # `python -m codex_claude_agent_mcp`
├── server.py # FastMCP server, 7 tools, core logic
├── config.py # env config + cwd/model/timeout validation
├── models.py # Pydantic tool input/result models
├── scheduler.py # per-process asyncio.Semaphore
├── claude_runner.py # ClaudeRunner ABC + RealClaudeRunner + FakeClaudeRunner
├── session_store.py # multi-process-safe SQLite store
├── errors.py # structured error codes + MCPError
└── logging.py # stderr logger with pid marker
The server depends only on the ClaudeRunner ABC, so tests use
FakeClaudeRunner and never call the real Claude API.
Every result carries an optional error field (null on success). Errors are
structured, never raw tracebacks:
{"code": "REVIEW_REQUIRES_ACCEPTANCE", "message": "...", "retryable": false, "details": {}}
Codes: INVALID_ARGUMENT, REVIEW_REQUIRES_ACCEPTANCE, INVALID_CWD,
CLAUDE_SESSION_ERROR, CLAUDE_AUTH_ERROR, CLAUDE_BILLING_ERROR,
CLAUDE_PROVIDER_ERROR, CLAUDE_PROTOCOL_ERROR, RATE_LIMITED, TASK_TIMEOUT,
SESSION_NOT_FOUND, SESSION_MISMATCH, JOB_CONTEXT_MISMATCH,
JOB_STATE_CONFLICT, CANCELLED, STATE_STORE_ERROR, INTERNAL_ERROR.
job_id is permanently bound to its original task, acceptance
criteria, resolved working directory, and (when supplied) its project_root.
Review rejects any mismatch; the root is not writable through the store.continue_task resumes only the execution session stored for that exact job
after Claude confirmed it, and inherits that job's project_root; it has no
root input of its own.project_root / ALLOWED_PROJECT_ROOTS are cwd/start-directory gates. They
are not an operating-system sandbox: execution Bash commands still run with the
MCP process account's permissions. Use narrow roots and run only on trusted code.The repository includes skills/codex-sol-claude-orchestrator/SKILL.md.
Copy that directory to ~/.codex/skills/ if you want Codex to apply the
delegation workflow automatically. It is intentionally separate from MCP
server installation.
Server 本身不绑定 Codex 可执行程序。其他 Agent 只要支持本地 STDIO MCP 和结构化
Tool Result,也可以使用;客户端必须为新 Job 提供可信 project_root,并负责安排
task、review 和 resume。更换主脑或执行端时见
AGENT_PORTING.md。
功能边界、统计口径、数据库迁移及验证方法见 v0.4 定稿。 后台任务不会跨 MCP 进程重启继续运行;重启后按确认过的 session 恢复,不能把 RUNNING 当成功。回归使用 Fake Runner / 模拟 SDK,不产生 Claude API 调用。
Licensed under the Xiaoye Noncommercial Reciprocal Source License 2.0: non-commercial use only, attribution required, and distributed or network-served Covered Works (including derivatives) must publish their Corresponding Source under the same license. Commercial use requires separate prior written permission.
Because commercial use is prohibited, this is a source-available license, not an OSI-approved open-source license.
The full bilingual terms, component boundaries, prior-version grants and source/data separation are documented in licensing.md and license history. An independent output is not automatically a derivative, but commercial use of the covered orchestration tool itself still needs permission. This custom license has not been reviewed by counsel and does not guarantee enforcement or compensation.
点将台:可迁移的 Agent 分层协作编排器。支持任务派发、会话续跑、独立审查、后台执行与用量记录。当前开箱即用为 Codex → Claude Code,可经适配迁移到其他 Agent 组合。
Tool
Belum ada daftar tool. Server remote mendapatkannya setelah pemeriksaan kesehatan pertama.
FAQ
Salin konfigurasi stdio untuk Cursor, Claude Desktop, Claude Code, atau VS Code di halaman ini. Konfigurasi ini memakai instalasi dari penerbit, bukan proxy PlugMCP.
Tidak ada variabel lingkungan wajib yang tercantum di catatan ini.
Codex Cc Orchestrator berjalan secara lokal melalui stdio.
Terakhir diperiksa 2026-10-28. 点将台:可迁移的 Agent 分层协作编排器。支持任务派发、会话续跑、独立审查、后台执行与用量记录。当前开箱即用为 Codex → Claude Code,可经适配迁移到其他 Agent 组合。