pi-mcp-audience

Filter MCP content based on audience annotations for Pi coding agent.

MCP resources and content blocks can carry annotations including an audience field indicating who the content is intended for — "user", "assistant", or both. This extension reads those annotations and hides content that is not meant for the user (e.g., raw JSON payloads, internal data, structured responses for the LLM only).

How it works

When an MCP tool returns content, each block may carry annotations:

{
  "type": "text",
  "text": "Here's what I found...",
  "annotations": { "audience": ["user", "assistant"] }
},
{
  "type": "text",
  "text": "{ \"raw\": \"json payload\" }",
  "annotations": { "audience": ["assistant"] }
}

• The first block includes "user" → displayed normally.
• The second block is assistant-only → hidden from the user display, but still visible to the AI (the LLM processes it, the user sees a placeholder).

Hidden items are replaced with a compact note:

[1 assistant-only content block hidden — visible to AI only]

Installation

1. Install pi-mcp-extension (with annotation support)

pi-mcp-audience works with the tools registered by pi-mcp-extension. It inspects the audience field in annotations on each content block, which requires annotation preservation in convertMcpContent().

A PR adding annotation preservation upstream is still in progress — until it merges, use the patched fork which already includes it:

npm install -g git+https://github.com/cwt/pi-mcp-extension.git

Once the upstream PR merges, annotation preservation will be built into pi-mcp-extension directly and pi install npm:pi-mcp-extension is all you need — no fork required.

2. Install pi-mcp-audience

pi install npm:pi-mcp-audience

Or for local testing:

pi -e src/index.ts

Usage

Toggle visibility

Method	Action
/mcp-audience	Toggle
/mcp-audience on	Show all content (including assistant-only)
/mcp-audience off	Hide assistant-only content (default)
Ctrl+Shift+A	Keyboard shortcut

Diagnostics

Command	Description
/mcp-audience debug	Show annotation stats for last N tool calls
/mcp-audience status	Current state summary

Footer status

3 item(s) hidden by audience filter

Compatibility

pi-mcp-extension

The primary target. Requires annotation preservation in convertMcpContent() — see installation step 1 above.

Any tool

Works with any tool returning annotations.audience — not just MCP tools. Harmless if no annotations are present (no-op).

Architecture

Two-phase approach:

Phase	Event	What happens	User sees	AI sees
1. Filter	tool_result	Replace assistant-only items with placeholders. Stash originals in details.__mcpAudienceOriginals	Placeholder	Placeholder (temporarily)
2. Restore	context	Before each LLM call, restore originals from details into the context copy (deep copy, safe to modify)	Unaffected	Full original data

The context event fires before every LLM call with a deep copy of messages. We restore the original content blocks there, so the AI always sees the complete data. The session's stored messages retain the filtered version for display.

Project Structure

pi-mcp-audience/
├── src/
│   ├── index.ts              # Extension entry point (default export)
│   ├── audience-filter.ts    # Core filtering logic
│   └── types.ts              # Shared type definitions
├── tests/
│   └── audience-filter.test.ts  # 12 tests
├── package.json
├── tsconfig.json
├── .gitignore
├── .npmignore
├── LICENSE
└── README.md

Technical Details

• Hook: tool_result (filter display) + context (restore for AI)
• Content inspection: Checks item.annotations.audience and item.resource.annotations.audience
• Persistence: Filtered indices + originals stored in details.__mcpAudienceOriginals
• Toggle state: In-memory (per session), affects future tool results
• No peer dependency: pi-mcp-audience does not import pi-mcp-extension. It hooks Pi's event system generically. If no annotations are present, it's a no-op.

Development

# Typecheck
npm run typecheck

# Run tests
npm test

# Test with pi
pi -e src/index.ts

License

MIT