Pi / oh-my-pi Integration

Plasmate integrates with both Pi (the original coding agent by @mariozechner) and oh-my-pi (the extended fork with MCP, LSP, browser, and subagent support) via MCP.

Both agents discover Plasmate's 13 tools automatically through their MCP configuration.

Quick Setup

oh-my-pi (recommended)

oh-my-pi has native MCP support via its mcp/ discovery layer. Add Plasmate to your MCP configuration:

{
  "servers": {
    "plasmate": {
      "command": "plasmate",
      "args": ["mcp"],
      "transport": "stdio"
    }
  }
}

Place this in your oh-my-pi MCP config file (typically ~/.config/oh-my-pi/mcp.json or passed via --mcp-config).

Once configured, oh-my-pi automatically discovers all 13 Plasmate tools:

Pi (original)

Pi supports MCP tools through its --mcp flag or configuration:

pi --mcp plasmate="plasmate mcp" "Research the pricing on stripe.com/pricing"

Or add to your Pi config file (~/.config/pi/config.toml):

[mcp.plasmate]
command = "plasmate"
args = ["mcp"]

Why Plasmate with Pi?

Pi and oh-my-pi both have built-in web browsing via their web/ module, which typically uses headless Chrome or simple HTTP fetch + readability extraction. Plasmate replaces this with structured SOM output that uses 4x fewer tokens.

Default Pi web browsing

User: What are the pricing tiers on stripe.com?

Pi uses web_search + fetch → raw HTML or markdown
→ ~30,000 tokens of page content in context
→ Model reasons over noisy input

With Plasmate

User: What are the pricing tiers on stripe.com?

Pi calls fetch_page via MCP → SOM JSON
→ ~8,000 tokens of structured content
→ Model receives typed regions, elements, and affordances

The token savings compound across multi-page research sessions. A 10-page research task that consumes 300,000 tokens with raw HTML uses approximately 80,000 tokens with SOM.

Multi-step Interaction Example

oh-my-pi's subagent and task orchestration layer works well with Plasmate's stateful session tools. A multi-step workflow:

1. open_page("https://example.com/login")
   → Returns SOM with form fields: email input (e3), password input (e5), submit button (e7)

2. type_text(session_id, "e3", "[email protected]")
   → Returns updated SOM

3. type_text(session_id, "e5", "password123")
   → Returns updated SOM

4. click(session_id, "e7")
   → Submits form, navigates to dashboard, returns new SOM

5. navigate_to(session_id, "https://example.com/settings")
   → Returns settings page SOM with all interactive elements

Each step returns the full updated SOM, so Pi always has current page state in context without re-fetching.

CDP Fallback

If you need pixel-perfect rendering or browser features that SOM does not cover (video playback, canvas interactions, WebRTC), Plasmate also exposes a CDP server:

plasmate serve --port 9222

Pi and oh-my-pi can connect via Puppeteer:

const browser = await puppeteer.connect({
  browserWSEndpoint: 'ws://127.0.0.1:9222'
});

Use MCP for structured browsing (research, data extraction, form filling). Use CDP when you need the full browser.

Authenticated Browsing

Plasmate supports cookie-based auth profiles for browsing sites the user is logged into. This works transparently with Pi:

# Set up an auth profile (one-time)
plasmate auth create twitter

# Pi automatically uses the profile when fetching authenticated pages
pi "What are the trending topics on my Twitter timeline?"
# → Plasmate serves authenticated SOM via the stored cookie profile

See the Authenticated Browsing guide for setup details.

Performance