Codex Supervisor MCP server
io.github.zriyox/codex-supervisor-mcp·v0.8.0·3 starsParallel Codex CLI workers from any MCP client: one git worktree each, state in SQLite, web board.
3 stars2.6k downloads/wk
Reviews
Write oneNobody has reviewed Codex Supervisor yet.
If you have run it, two minutes of your experience saves the next person an afternoon.
Codex Supervisor tools (19, 4 write)
write = sends, deletes, buys or postsRead from the package source without running it. The installed server may list more.
ask_codex_workercancel_codex_workerwrite actionTerminate a running Codex worker. Uses the in-process handle when available, otherwise the pid recorded on disk so a different MCP session can still cancel it.
check_for_updatewrite actionAsk the npm registry whether a newer codex-supervisor-mcp is published, and whether the installed files match the published tarball. Returns installed and latest version, integrity check, and the install command. The same check runs on startup; call this to force a fresh one.
create_codex_followup_workerwrite actionStart a NEW Codex session seeded with the previous worker's prompt, status, and recent events. To continue the same session instead, use resume_codex_worker.
create_codex_workerwrite actiondescribe_sessionRecord what a session is about: a one-line title and an optional note. Readers of the store (the web view, get_session_works) show these instead of the bare session id. Fields you omit are kept.
get_codex_worker_eventsRead recent raw Codex JSONL events for one worker. limit, maxChars and kinds keep the read inside a context budget; maxChars truncates the middle of oversized strings the same way Codex truncates its own output.
get_codex_worker_statusRead the latest normalized status for one Codex worker: lifecycle, phase, current action, changed files, goal, recent commands, and the task prompt (clipped by default). Set includePrompt to read the prompt in full, or use get_codex_worker_events for the raw stream.
get_orchestration_overviewCompact status table for every worker, budgeted to stay under ~2k tokens. This is the default read for an orchestrating main thread; fetch details per worker only when needed.
get_session_worksEvery worker dispatched under one session_id, oldest first. This is the reconnect path: a main-thread session that died can list the whole batch it started without holding any of it in context.
get_worker_diffThe worker's actual changes as patches: everything between the commit its worktree started from and its working tree, committed or not, plus untracked files. This is the ground truth to check a worker's report against. maxChars bounds the whole answer: files past the budget come back listed but without a patch, so the file list is always complete. paths narrows the diff to those files or directori
get_worker_goalRead the goal recorded for a worker, plus the native Codex thread goal when the thread has one (status, tokens used, time used).
get_worker_resultget_worker_summaryOne compact summary per worker: goal, status, changed files, last command, last message. Cheap enough for a main thread to read every worker at the end of a run.
land_codex_workerlist_codex_workersList tracked Codex workers. Defaults to active workers only.
resume_codex_workersearch_worksSubstring search over title, goal, prompt and last message of every tracked worker, newest first. Use it to find an earlier work by keyword instead of paging through history.
wait_codex_workers
Public scan report
scanner v0.1.10 · 2026-10-13 · same rubric, same numbers if you re-run it
- Code scan24 source files scanned15/25
- –Live reliabilityno gateway calls yet and no remote to proben/a
- –Tool poisoningtools not inspected (local package is not executed); not countedn/a
- Auth qualitylocal package, no credentials required12/15
- Maintenancelast push 0 days ago15/15
- Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10
Findings (2)
- mediumnpm install lifecycle script present
install.script - mediumnpm install lifecycle script present
install.scriptpackage.json: …, "LICENSE" ], "scripts": { "postinstall": "node ./src/setup.js --auto --quiet", …
How Codex Supervisor compares
Against every graded Databases MCP server, from our own scans and uptime checks. Updated after each scan run.
- Safety score
- 77/100
- higher than 40% of 803 Databases servers
- Tools
- 19
- the typical Databases server has 8
- Tools that change things
- 21%
- typical: 10%. Fewer write tools means less an agent can break
- GitHub stars
- 3
- more than 65% of 601 Databases servers
- Downloads a week
- 2,631
- more than 95% of 315 Databases servers with a package
What its tools let an agent do: 15 read · 2 create · 1 change · 1 delete
Install Codex Supervisor in Claude Code, Cursor or VS Code
Runs npx -y codex-supervisor-mcp on your machine. Read the scan report first; the gateway never runs local packages.
claude mcp add codex-supervisor-mcp -- npx -y codex-supervisor-mcp
What the publisher says
From the Codex Supervisor repository's README, as published. We do not edit it. Read it on GitHub
codex-supervisor-mcp
English | 中文
官网 codex-supervisor.zriyo.com · 给模型读的 llms.txt
让几个 agent 同时改一个仓库,结果是互相覆盖、没人说得清谁改了什么,主线程的上下文还被 worker 的输出塞满。
codex-supervisor-mcp 是一个 Codex MCP server,把「派单」和「记账」从模型上下文里拿出来放到磁盘上:一个主线程(Claude Code、Codex 或任何 MCP 客户端)同时指挥多个 Codex CLI worker,一个 worker 一个 Git worktree,状态落盘到 SQLite,附一个网页看板。
30 秒真跑:createcodexworker ×3,三个 worker 各自 worktree 并行,主线程收 3 份 diff 合成一次 commit。
解决什么
worker 跑的是 codex exec,model 透传:主线程留在 Claude,worker 可以挂 DeepSeek 或任何 Codex 配了 provider 的模型,账单分开算。
五分钟跑起来
前置:Node.js 22.13.0 以上,codex CLI 在 PATH 里。macOS、Linux、Windows 都行。
1. 装
npm install -g codex-supervisor-mcppostinstall 装 skill、注册 MCP。claude mcp list 里没看到(npx、pnpm、--ignore-scripts 不跑 postinstall)就手动补:
claude mcp add -s user codex-supervisor -- npx -y codex-supervisor-mcp
codex mcp add codex-supervisor -- npx -y codex-supervisor-mcp重启 Claude Code / Codex。以后不用再手动更,见「更新」。只要 skill 不要 MCP:npx skills add zriyox/codex-supervisor-mcp。
2. 派第一批活
在 Claude Code 里直接说人话,skill 会让它走正确的流程:
把这三个模块的单测补齐,用 codex-supervisor 分三路并行跑,session 叫「补单测」。
主线程背后做的事(你也可以自己调工具):
create_codex_worker ×3 每路带 session_id、session_title、ownedPaths、goal
wait_codex_workers 默认等 2 分钟,到点返回快照,没完就接着等
get_worker_result ×3 收每路的汇报,带它真跑过的命令和退出码每路的改动在各自的 codex/ 分支上,合不合、怎么合由主线程定。
3. 开看板
codex-supervisor-web
# 没全局装也能起:
npx -p codex-supervisor-mcp codex-supervisor-web打开 http://127.0.0.1:7877。状态目录按 SUPERVISORHOME、最近的 .mcp.json、~/.codex-supervisor 的顺序找;端口用 SUPERVISORWEB_PORT 改。看板只看,不派单不取消。
我自己怎么用
能连 MCP 的都能当主线程,这里只是我的用法。我开两个 Claude Code 会话,一个只管文档,一个只管派活:一个会话又写详设又盯 worker,上下文两小时就满。
两个会话之间只传一段文字,粘进主脑会话用 /goal 接上。结构固定:
Shortened. The full README is on GitHub.
Nothing above is checked by us. What we check is on the safety report.
Codex Supervisor: common questions
- Is Codex Supervisor MCP server safe?
- Mostly: it is graded B (77/100). Read the Codex Supervisor safety report
- How do I install Codex Supervisor?
- It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.
- Does Codex Supervisor need an API key?
- No secret keys are declared. It reads 2 settings from the environment.
- Is Codex Supervisor maintained?
- The last commit was 19 days ago (2026-10-13). The latest release is v0.8.0.
- What can I use instead of Codex Supervisor?
- Servers from other publishers that do the same job: Remem MCP server, DBHub MCP server and Mindwtr MCP server.
- How does Codex Supervisor compare with other Databases MCP servers?
- safety score 77/100 (higher than 40% of 803 Databases servers); tools 19 (the typical Databases server has 8); tools that change things 21% (typical: 10%. Fewer write tools means less an agent can break). Full comparison.
Alternatives to Codex Supervisor
Same job from other publishers: the closest match first, then the best rated.
- RememLocal-first automatic memory for Claude Code and Codex: capture, distill, recall from SQLite.not reviewedEstablishedB
- DBHubToken-efficient database MCP server for PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, SQLitenot reviewedWidely usedA
MindwtrTask automation via SQLite, desktop Local API, or Mindwtr Cloud; read-only by default.not reviewedEstablishedA- Sequential Thinking Working NotesLocal working notes, decisions, evidence and resumable sessions with SQLite persistence.not reviewedEstablishedA
Sling CLISling CLI MCP server for querying any database, running data pipelines and managing replicationsnot reviewedEstablishedB