PG Memory Bank

Governance, security checks and TapSign integration for your editor — and the command line, on macOS, Linux and Windows.

Installing this extension is the whole setup. The command-line tools and the TapSign MCP server ship inside it. Nothing to clone, no account, no payment.

Start here

1. Add the tools to your PATH (once per machine)

macOS / Linux — add to ~/.zshrc or ~/.bashrc:

export PATH="$HOME/.pg-cli:$PATH"

Windows — add to your PATH:

%USERPROFILE%\.pg-cli

Open a new terminal, then check it worked:

pg status

You should see Session status: .... If you get "command not found", the PATH line has not been picked up yet — open a new terminal.

2. Set up a project

cd your-project
pg install backend --target .

That scaffolds the memory bank, the governance hooks and the agent instructions into your project.

3. Before you start coding, each day

pg start -Yes

4. Before you commit

pg self-check -EnableDbIndexMaintenanceCheck

The commands you will actually use

Working on an older project

A codebase that predates PG will not have a memory bank. Map it first — this reads your existing code and database schema and writes what it finds:

pg map-structure
pg install backend --target .
pg start -Yes

TapSign integration

pg tapsign-catalog                                                    # what can be protected
pg tapsign-plan -Profile universal-web -Endpoint POST:/login          # plan it
pg tapsign-integrate -Profile wordpress -Endpoint POST:/pay           # apply it
pg tapsign-verify -Profile guardian -Endpoint POST:/admin/delete-user # prove it

In the editor

Press Ctrl+Shift+P (Cmd+Shift+P on Mac) and type Narrate:

Tell your AI assistant what to do

You do not need to know how to write the instruction. In the Start Here panel, click Copy a Prompt for Your AI Assistant, pick what you want, and paste it into Copilot Chat, Claude or Codex.

Use Agent mode, not Ask. Tools only work in Agent mode.

The same prompt works in all three. Here it is if you would rather copy from here:

Use the TapSign MCP tools to integrate TapSign into this project.

Work in this order and do not skip ahead:

1. Call tapsign_start with the absolute path to this repository.
   Report exactly what it found: stack, whether consume is called,
   and the number of blocking problems.

2. If there are ANY blocking problems, call tapsign_verify and fix
   every one before doing anything else. A bypass makes every other
   control decorative — do not start new work while one exists.

3. Call tapsign_scenarios and tell me which apply here. Then ASK ME
   to choose before you implement anything. Do not assume.

4. For the scenario I pick, call tapsign_examples and tapsign_template,
   then show me the exact changes you propose. Wait for my approval.

5. Call tapsign_config and tell me which config files I need to
   download from the portal and exactly where each one goes.

Rules you must follow:
- Never write code that mints or verifies its own approval proof.
  Approval is proved only by calling a TapSign consume endpoint.
- Never put the connector key in anything the client can download.
- If TapSign is unreachable, the protected action must be REFUSED.
- Do not tell me integration is done based on a static scan. It cannot
  observe behaviour.

Start with step 1 and stop after reporting.

Two more are in the panel: audit an existing integration (read-only, tells you whether it can be bypassed) and run the checks and explain them.

No editor? Run the tools directly

node ~/.pg-cli/mcp/dist/index.js

Speaks MCP over stdio, so any MCP client can use it — including claude in a terminal, which reads the same .vscode/mcp.json.

The TapSign MCP server

Ships with the extension and installs to ~/.pg-cli/mcp. To use it in a project:

1. Ctrl+Shift+P → Narrate: Set Up TapSign MCP Server
2. Reload the window
3. Ask your AI assistant to run tapsign_start

Ten tools: tapsign_start, tapsign_verify, tapsign_diagnose, tapsign_config, tapsign_rules, tapsign_capabilities, tapsign_examples, tapsign_faq, tapsign_scenarios, tapsign_template.

The config is written to .vscode/mcp.json in that project only. Granting a server access to a codebase is your decision, so it is never attached globally.

What it checks

• Security — injection, XSS, SSRF, weak crypto, authentication and authorisation gaps
• Dependencies — known vulnerabilities; a high-severity advisory blocks
• Frontend standards — keyboard covering inputs, camera and scanner handling, spacing and headers, file and function size limits
• Governance — the memory bank must stay in step with the code
• Secrets — no credential can be committed. This one has no bypass.

Requirements

• PowerShell 7 (pwsh) — the tools are PowerShell, on every platform
• Python 3 — used by several guards
• Node.js — for the MCP server

pg doctor tells you which of these are missing.

Free

Everything is free. No account, no plan, no payment to install or use it.

If paid plans are ever introduced, anyone already using the extension keeps full access permanently. The date your machine first ran it is stored locally and never leaves your computer.

Security scanning, secrets guarding and governance checks will never depend on a licence. Those protect you; they are not a feature to be sold back to you.

Problems

Use the Q&A tab on the Marketplace listing, or narrate.pgdigitalglobal.com.

If a command misbehaves, pg doctor and Narrate: Run Command Diagnostics usually name the cause.