Codex Subagent
Delegate coding tasks to the OpenAI Codex CLI, installed separately, with explicit model and effort.
Open the source (opens github.com) https://github.com/parisbs/codex-subagent-mcp
Before you use it
- Runs on your computer with your permissions. Read what it can do before connecting it to an agent.
What it is
MCP servers from io.github.parisbs.
The problem it solves and who it suits haven't been reviewed by InnerVault. The description above is the source's own.
How to get it
npm package (stdio)
codex-subagent-mcp@0.6.0
From the server's entry in the official MCP Registry
Settings it asks for
CODEX_SUBAGENT_MAX_DELEGATIONS_PER_HOUR· Delegation and follow-up processes per rolling hour: decimal integer 1 to 9007199254740991; unset means unrestricted. In memory per server; probes never count.CODEX_SUBAGENT_MAX_BACKGROUND_JOBS· Cap on running background jobs: decimal integer 1 to 8 (default 8). Blocking calls do not use this cap.CODEX_SUBAGENT_USAGE_LOG· Opt-in local usage log: on or off (default off). Records counts and settings, never prompts, code or paths; labels are caller text.CODEX_SUBAGENT_DEFAULT_MODEL· Model used when a call names none. Unset: the only entry of a single-entry CODEX_SUBAGENT_ALLOWED_MODELS, otherwise a call without a model is refused with a…CODEX_SUBAGENT_DEFAULT_EFFORT· Reasoning effort used when a call specifies none: none, minimal, low, medium, high, xhigh, max or ultra. Unset: the model's own default. Cannot exceed…CODEX_SUBAGENT_ALLOWED_MODELS· Comma-separated allow-list of model slugs; anything else is refused. Unset: unrestricted. A single entry acts as the default model; an empty entry is a…CODEX_SUBAGENT_DEFAULT_SANDBOX· Sandbox used when a call specifies none: read-only, workspace-write or danger-full-access. Defaults to read-only and cannot exceed the ceiling.CODEX_SUBAGENT_MAX_SANDBOX· Ceiling on what a delegation may do: read-only, workspace-write or danger-full-access. Defaults to workspace-write; danger-full-access must be set explicitly.CODEX_SUBAGENT_MAX_EFFORT· Ceiling on reasoning effort: none, minimal, low, medium, high, xhigh, max or ultra. Unset: unrestricted. Higher efforts are lowered to a supported level, or…CODEX_SUBAGENT_MCP_SERVERS· MCP servers from Codex's config a delegation keeps: none (default), all, or comma-separated names.CODEX_SUBAGENT_PLUGINS· Installed Codex plugins a delegation keeps: none (default), all, or comma-separated name@marketplace ids.CODEX_SUBAGENT_APPS· Whether a delegation keeps Codex's apps (connectors to external services): off (default) or on.
Related
-
my-codex
MCP serversAll-in-one agent harness for OpenAI Codex CLI — Boss meta-orchestrator, 17 curated agents (+17 opt-in AI specialists), 123 curated skills, 3 MCP servers. Install once, curated specialists ready.
git clone https://github.com/sehoon787/my-codex.git -
deepseek-as-subagent
MCP serversRun DeepSeek as a real sub-agent inside Claude Code / Codex CLI — with its own 7-tool agent loop, workspace-scoped file access, and bounded Bash.
git clone https://github.com/PsChina/deepseek-as-subagent.git -
Subagent MCP
MCP serversLet Codex orchestrate external coding agents through their native harnesses.
subagent-harness-mcp@1.0.31 -
agent-swarm
MCP serversYour Company Agentic Operating System
git clone https://github.com/desplega-ai/agent-swarm.git -
sub-agents-mcp
MCP serversDefine task-specific AI sub-agents in Markdown for any MCP-compatible tool.
git clone https://github.com/shinpr/sub-agents-mcp.git -
YHWH
MCP serversKether-governed AI agent orchestration with Pi/MCP, isolated workers, LSP tools, and a portable Windows/WSL2 installer
git clone https://github.com/Yuzu482/YHWH.git