Settings — MCP Connection¶

Last updated: 2026-09 · owner: Youngjin · volatility: low

L0 TL;DR: This setup lets your AI tool (Claude Code, Codex, Kiro, Amazon Quick) read Playbook pages, the Radar, and evidence records through tool calls. Download one MCP¹ server file and register it; that is all. The content is public, so there is no login or API key.

What gets connected¶

Item	Value
Server file	`https://comeddy.github.io/pai-playbook/mcp/pai-playbook-mcp.mjs` (single file, no npm install)
Requirements	Node.js 18+ · macOS/Linux terminal (WSL on Windows) · internet access (reads content from GitHub raw)
Transport	Local · stdio² (default) · remote HTTP not configured
Content	This site's markdown sources (four languages) and `assets/claims.json` — identical to the site, read-only

Four tools are provided:

Tool	Purpose
`playbook_list_pages`	List of page ids, titles, and URLs
`playbook_read_page`	Source of one page (`page`, `lang`)
`playbook_search`	Search a string across all pages → snippets
`playbook_evidence`	Evidence-record summary or one claim in detail

Per-client procedure¶

Claude CodeCodexKiro CrewAmazon QuickKiro CLI

Claude Code runs the local MCP server. If a registration already exists, only update its connection details.

① Download the server file

mkdir -p "$HOME/.pai-playbook" && curl --fail --location --show-error \
  --output "$HOME/.pai-playbook/server.mjs" 'https://comeddy.github.io/pai-playbook/mcp/pai-playbook-mcp.mjs'

② Register in Claude Code — registers pai-playbook at user scope. Model and tool-approval settings are not changed.

claude mcp add pai-playbook --transport stdio --scope user \
  -- node "$HOME/.pai-playbook/server.mjs"

③ Verify the connection — check in /mcp that pai-playbook is connected, then ask for a real lookup.

/mcp

Tell me the latest Physical AI news from the Playbook

Remote · HTTP — address not configured

No public HTTP MCP address is registered for this site yet. The steps below are for reference; replace the placeholders (YOUR_…) with real values before use. They assume a public OAuth client with the callback http://localhost:9876/callback registered.

claude mcp add --transport http --scope user \
  --client-id 'YOUR_CLIENT_ID' --callback-port 9876 \
  pai-playbook 'YOUR_HTTP_MCP_URL'

Claude Code official MCP docs ↗

How to register a local MCP server in the Codex CLI. Commands follow the official docs and have not been verified on a real device in this repo ([4]).

① Download the server file

mkdir -p "$HOME/.pai-playbook" && curl --fail --location --show-error \
  --output "$HOME/.pai-playbook/server.mjs" 'https://comeddy.github.io/pai-playbook/mcp/pai-playbook-mcp.mjs'

② Register in Codex

codex mcp add pai-playbook -- node "$HOME/.pai-playbook/server.mjs"

③ Verify the connection — reopen Codex and ask for a lookup.

Search the Playbook for 'HyperPod'

Remote · HTTP — address not configured

Replace the placeholders with real values. Send the callback URL the command prints to the administrator for registration, then log in with codex mcp login pai-playbook.

codex mcp add pai-playbook --url 'YOUR_HTTP_MCP_URL' \
  --oauth-client-id 'YOUR_CLIENT_ID'

Codex MCP docs ↗

The server file must be on the machine where Crew runs. The steps follow the official docs and have not been verified on a real device ([4]).

① Download the server file

mkdir -p "$HOME/.pai-playbook" && curl --fail --location --show-error \
  --output "$HOME/.pai-playbook/server.mjs" 'https://comeddy.github.io/pai-playbook/mcp/pai-playbook-mcp.mjs'

② Add a local server in Crew — in Agent Capabilities → Integrations (MCP), add a command-based server, or merge the entry below into the MCP config Crew reads. Replace the path placeholder with the real absolute path.

printf '%s\n' "$HOME/.pai-playbook/server.mjs"

{
  "mcpServers": {
    "pai-playbook": {
      "command": "node",
      "args": ["ABSOLUTE_PATH_TO_SERVER_MJS"]
    }
  }
}

③ Discover and verify — apply the config with Discover & Sync, confirm the connection with Probe All, then ask for a lookup.

Show me the latest intake items on the Playbook Radar

Remote · HTTP — address not configured

There is no address yet for adding a URL-based server in the Crew dashboard. Once one is registered, enter YOUR_HTTP_MCP_URL as the server URL. This server serves public content, so no authentication header is needed.

Kiro Crew MCP docs ↗

Amazon Quick supports remote HTTP MCP only. This site has no public HTTP address yet, so it cannot be connected at the moment. Once an address is registered, proceed as follows ([4]).

In Connectors → Create for your team → Model Context Protocol (MCP), create a new connection and enter the server URL YOUR_HTTP_MCP_URL.
Authentication follows the server-side configuration. For a public-content server, choose no authentication.
Use the connector's Sync to load the tool list, then ask for a lookup. Quick's task timeout is 60 seconds.

Summarize the Cloud vs Edge decision criteria from the Playbook decision trees page

Amazon Quick MCP docs ↗

Registers a local MCP server on the machine where the Kiro CLI runs. The configuration follows the official docs and has not been verified on a real device ([4]).

① Download the server file

mkdir -p "$HOME/.pai-playbook" && curl --fail --location --show-error \
  --output "$HOME/.pai-playbook/server.mjs" 'https://comeddy.github.io/pai-playbook/mcp/pai-playbook-mcp.mjs'

② Add to the Kiro CLI settings — merge the entry below into mcpServers in ~/.kiro/settings/mcp.json. Keep existing servers and approval settings, and replace the path placeholder with the real absolute path.

printf '%s\n' "$HOME/.pai-playbook/server.mjs"

{
  "mcpServers": {
    "pai-playbook": {
      "command": "node",
      "args": ["ABSOLUTE_PATH_TO_SERVER_MJS"]
    }
  }
}

③ Verify the connection

/mcp

Summarize the P4 Sim-to-Real page from the Playbook

Remote · HTTP — address not configured

Replace the placeholders with real values. The callback http://localhost:9876/callback must be registered; use /mcp auth to re-authenticate.

{
  "mcpServers": {
    "pai-playbook": {
      "url": "YOUR_HTTP_MCP_URL",
      "oauth": {
        "clientId": "YOUR_CLIENT_ID",
        "redirectUri": "http://localhost:9876/callback",
        "oauthScopes": ["openid", "email", "profile"]
      }
    }
  }
}

Kiro CLI MCP docs ↗

After connecting, ask like this¶

Tell me the latest Physical AI news from the Playbook

Put the latest intake items on the Playbook Radar into a table

Put the Build vs Buy decision criteria from the Playbook decision trees into a table

Verification scope and limits¶

The Claude Code procedure was confirmed on 2026-09-26 by registering this repo's server file at user scope and checking /mcp connected plus a page-list lookup. The Codex, Kiro Crew, Amazon Quick, and Kiro CLI commands follow their official docs and are unverified ([4]).
The server reads markdown from the main branch. Committed content may appear before the site reflects it; the cache lasts 10 minutes.
Until the evidence records (assets/claims.json) are published on the site, playbook_evidence returns a "not published yet" notice.
The server file has no write tools and handles no credentials. For offline use, clone the repo and set the PAI_PLAYBOOK_DOCS_DIR=<clone>/docs environment variable.
Report problems and suggestions as GitHub issues.

➡️ Next action: Register the one client you use most and try playbook_search with one keyword before your next customer meeting.

owner: Youngjin · updated: 2026-09 · volatility: low

MCP (Model Context Protocol) — An open protocol that lets AI tools call external data and functions in a standard way. A server exposes "tools" and a client such as Claude Code invokes them. ↩
stdio transport — An MCP connection mode in which the client runs the server program on the same machine and talks to it over standard input/output. No network address or login is needed. ↩