CURRENT CLI KNOWLEDGE / READ ONLY

Help your coding agent master its own CLI.

AI coding CLIs change faster than model training data. Connect this read-only MCP so your LLM can check source-backed commands, configuration, models, effort levels, permissions, and workflows before it advises or acts.

supported coding CLIs
3
localized source-checked manuals
162
structured read-only tools
12
01 / UNDERSTAND

What your LLM can learn before acting

Use one CLI or all three. Retrieval stays provider-specific, so similar-looking flags and settings are never silently translated across products.

01

Exact CLI grammar

Look up commands, flags, configuration, environment variables, workflows, and risk notes as structured records.

02

Provider-specific meanings

Keep Claude Code, Codex, and Grok Build syntax separate—including the three different meanings of -p.

03

Model and effort choices

Read captured model catalogs, supported native effort levels, availability warnings, fallbacks, and live-verification steps.

04

Current-CLI work plans

Create bounded workstreams around the CLI you already use, with dependencies, permission boundaries, and verification gates.

05

A deeper path for one hard problem

Ask for the provider-local maximum-depth candidate and supported single-task effort without inventing a universal winner.

06

Evidence you can inspect

See captured release, review date, canonical link, official sources, and any runtime/documentation drift behind the answer.

03 / CONNECT

Connect the CLI you actually use

You do not need all three clients. Add the endpoint to your current CLI, verify the registration, then let that LLM query its own operating reference.

02

Codex

Codex-native guidance for profiles, codex exec, sandboxing, models, and reasoning effort.

Add server
codex mcp add worlds-knowledge --url https://mcp.claude-world.com/mcp
Verify connection
codex mcp get worlds-knowledge --json
03

Grok Build

Grok-native guidance for single prompts, models, sandboxing, permissions, and workflows.

Add server
grok mcp add --transport http --scope user worlds-knowledge https://mcp.claude-world.com/mcp
Verify connection
grok mcp doctor worlds-knowledge --json
04 / ASK

Three useful ways to invoke it

Start with the current-CLI check. Advisor and planner are optional deeper paths, not requirements for using the MCP.

01 / PRIMARY

Make Claude Code check itself first

This is the primary job: retrieve the current CLI reference before suggesting syntax or a workflow.

Before acting, use AI Worlds Knowledge MCP to inspect the source-checked Claude Code reference for this task. Show the exact non-interactive usage, model selection, supported effort, and permission mode; cite the returned sources, include the captured version and review date, and state what I must verify against my installed claude version.
02 / ADVISOR

Route one hard problem deeper

Ask for the maximum-depth role inside the active provider, then verify the snapshot against the account-visible model catalog.

Call worlds_advise_hard_problem first for this single difficult problem. My active and only available CLI is claude-code; use selection_scope current-cli and keep execution_intent read-only unless I explicitly authorize edits. Then verify the recommendation with worlds_list_cli_models and my live catalog. Show the provider-local maximum-depth candidate, supported single-task effort, fallback, POSIX template, and evidence without widening permissions or enabling orchestration.
03 / PLANNER

Plan work by model, effort, and boundary

For larger tasks, request an advisory split that keeps your current CLI in control unless another installed, authorized CLI has a concrete advantage.

Use worlds_list_cli_models and worlds_plan_cli_work to plan this task around my current CLI. Keep audit and review workstreams read-only; mark only implementation I have already authorized as may-edit. For each workstream show the concrete model, supported native effort, dependencies, permission boundary, fallback, POSIX command template, and verification gate. Do not launch another CLI or grant permission.
05 / 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 exposes no local-file, local-setting-value, write, command-execution, or arbitrary-URL-fetch tool.
  • may-edit describes an already-authorized plan; it never gives the MCP write access or expands your agent permissions.
  • Returned commands are POSIX reference templates. Your agent, shell, installed version, and approval policy still decide whether anything runs.
06 / FRESHNESS

Freshness stays visible

  • Results preserve product, captured release, source status, review date, canonical link, and official sources.
  • Model recommendations fail closed when the validated snapshot is stale; your live account catalog remains authoritative for availability and entitlement.
  • Runtime, documentation, and source disagreements remain visible as drift. Recheck consequential guidance against the returned primary source.
  • The public corpus is 54 CLI modules localized into three languages—not website news, opinion, API companions, or release archives.