GPT WORLD / INDEPENDENT CODEX FIELD MANUAL Source review · 2026-07-31
GPT World
Open navigation
EN 中 日

CURRENT CODEX OPERATING LAYER / READ ONLY

Give your LLM a current Codex operating manual.

Model training ages; the Codex CLI keeps moving. Worlds Knowledge MCP lets your LLM check versioned commands, model and effort support, permission boundaries, and official sources before it tells you what to run.

Connect Codex
released CLI snapshots
3
source-checked manuals
162
read-only MCP tools
12
Hosted endpoint 
https://mcp.claude-world.com/mcp
01

What your LLM learns before you install anything

Six bounded capabilities turn remembered Codex advice into a current, inspectable operating decision.

01

Exact Codex syntax

Look up codex exec, -p profiles, flags, config keys, environment names, and workflows from a captured release.

02

Model and effort fit

Keep model selection, supported reasoning effort, orchestration, sandbox, and approval policy as separate choices.

03

Hard-problem advisor

Route one difficult problem to the provider-local maximum-depth catalog role, with a supported effort and fallback.

04

Workstream planning

Decompose bounded tasks with ownership, dependencies, permission boundaries, handoffs, and verification gates.

05

Freshness and drift

Expose capture dates and source drift, fail closed on stale model snapshots, and defer to the live account catalog.

06

Provider-safe teaching

Teach from 162 localized, source-checked manuals without translating Claude Code or Grok Build flags into Codex.

02

Connect the one client you actually use

Codex is the recommended path on GPT World. You do not need all three CLIs: install only the card for your active client, then run its check.

Codex

Recommended here
Install 
codex mcp add worlds-knowledge --url https://mcp.claude-world.com/mcp
Verify 
codex mcp get worlds-knowledge --json

Claude Code

Alternative client
Install 
claude mcp add --transport http --scope user worlds-knowledge https://mcp.claude-world.com/mcp
Verify 
claude mcp get worlds-knowledge

Grok Build

Alternative client
Install 
grok mcp add --transport http --scope user worlds-knowledge https://mcp.claude-world.com/mcp
Verify 
grok mcp doctor worlds-knowledge --json
03

Ask for current Codex facts first

This prompt forces exact reference lookup and separates model, effort, permission, and execution instead of trusting model memory.

PROMPT / 03
Use the Worlds Knowledge MCP to verify current Codex exact syntax, model, effort, and permission behavior. Call worlds_list_cli_versions, worlds_search_cli_reference, the exact command or configuration tool, and worlds_list_cli_models. Explain what -p means, which command is non-interactive, how to select a model and supported reasoning effort, and how sandbox and approval policy differ. Treat my live account catalog as authoritative, cite returned official sources, and do not execute anything.
04

Give one hard problem to the deepest eligible Codex model

The advisor chooses the provider-local maximum-depth catalog role for one bounded problem. It does not enable orchestration or claim a universal winner.

PROMPT / 04
Call worlds_advise_hard_problem first for this one difficult problem with current_cli codex, selection_scope current-cli, and execution_intent read-only. Then verify the recommendation with worlds_list_cli_models and my live account catalog. Return the supported single-task effort, fallback, entitlement uncertainty, and a source-checked POSIX command template. Do not use ultra orchestration, launch Codex, or widen permissions.
05

Plan Codex work by model, effort, and boundary

Use the planner for multiple workstreams; it keeps the active CLI, execution authority, effort, orchestration, and verification separate.

PROMPT / 05
Use worlds_list_cli_models and worlds_plan_cli_work to plan this repository task. My current and only available CLI is codex. Decompose audit, implementation, and final review into bounded workstreams with dependencies and verification. Keep audit and review read-only; use execution_intent may-edit only for implementation I have explicitly authorized. Show each concrete model, supported effort, fallback, POSIX template, sandbox and approval boundary. Do not launch Codex, enable ultra orchestration, or grant permissions.

ROADMAP / NEXT KNOWLEDGE LAYER

System operations come next—not as guessed shell snippets.

Version 0.5.0 does not include a general system-operations corpus. A future layer can teach the computer-operation knowledge an LLM needs while preserving platform and risk boundaries.

  1. 01Shell and filesystem
  2. 02Git and version control
  3. 03Process and network
  4. 04Package and build tools
  5. 05Containers and cloud CLIs
  6. 06macOS, Linux, and Windows

Future entries must be platform-specific, source-checked, and risk-labeled. POSIX shell, PowerShell, and cmd.exe syntax must never be presented as interchangeable.

07 / BOUNDARY

Read-only by design

  • Your query text and selected filters are sent to the hosted endpoint. Do not include secrets, credentials, or private code.
  • The MCP has no local-file, local-setting-value, write, command-execution, or arbitrary-URL-fetch tool.
  • execution_intent may-edit changes only an advisory template; it never grants the MCP or your LLM write authority.
  • Generated command templates are labeled POSIX. Adapt and recheck them before using PowerShell or cmd.exe.
08 / SOURCES

Freshness stays visible

  • Released runtime evidence is preferred, then official documentation; unreleased source evidence remains labeled preview.
  • Model catalogs are dated snapshots with a seven-day freshness gate. Your live account catalog is always authoritative.
  • Results retain product, release, source status, review date, canonical link, and official sources so drift remains inspectable.
  • The public MCP returns bounded manual summaries and canonical teaching links; website editorial content is not packaged into it.