Pi MCP Adapter

Use MCP servers with Pi without burning your context window.

https://github.com/user-attachments/assets/4b7c66ff-e27e-4639-b195-22c3db406a5a

Why This Exists

Mario wrote about why you might not need MCP. The problem: tool definitions are verbose. A single MCP server can burn 10k+ tokens, and you're paying that cost whether you use those tools or not. Connect a few servers and you've burned half your context window before the conversation starts.

His take: skip MCP entirely, write simple CLI tools instead.

But the MCP ecosystem has useful stuff - databases, browsers, APIs. This adapter gives you access without the bloat. One proxy tool (~200 tokens) instead of hundreds. The agent discovers what it needs on-demand. Servers only start when you actually use them.

Install

pi install npm:pi-mcp-adapter

Restart Pi after installation.

What happens on first run

The adapter reads standard MCP files automatically. No extra setup needed if you already have them.

| You already have... | What happens | |---------------------|--------------| | .mcp.json or ~/.config/mcp/mcp.json | Pi uses it immediately. The first time you open /mcp, you'll see a short heads-up explaining which file Pi detected and that Pi only writes adapter-specific overrides to its own files. | | Host-specific configs (Cursor, Claude Code, Codex, etc.) but no standard MCP files | Run /mcp setup to adopt those host configs into Pi. The setup flow shows exactly what it found, lets you pick which ones to import, and previews the exact file changes before writing. | | Nothing configured yet | Run /mcp setup to scaffold a minimal .mcp.json, quick-add RepoPrompt, or inspect what the adapter discovered on your machine. |

If you prefer the terminal, you can also run pi-mcp-adapter init after install to scan for host-specific configs and add missing compatibility imports to the Pi agent dir (~/.pi/agent/mcp.json by default, or $PI_CODING_AGENT_DIR/mcp.json when set).

Quick Start

Preferred project config: .mcp.json

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    }
  }
}

Preferred user-global shared config: ~/.config/mcp/mcp.json

Pi also reads Pi-owned override files for settings and host-specific compatibility:

• <Pi agent dir>/mcp.json — Pi global override (~/.pi/agent/mcp.json by default)
• .pi/mcp.json — Pi project override

Precedence is:

1. ~/.config/mcp/mcp.json
2. <Pi agent dir>/mcp.json
3. .mcp.json
4. .pi/mcp.json

Servers are lazy by default — they won't connect until you actually call one of their tools. The adapter caches tool metadata so search and describe work without live connections.

mcp({ search: "screenshot" })

chrome_devtools_take_screenshot
  Take a screenshot of the page or element.

  Parameters:
    format (enum: "png", "jpeg", "webp") [default: "png"]
    fullPage (boolean) - Full page instead of viewport

mcp({ tool: "chrome_devtools_take_screenshot", args: '{"format": "png"}' })

Note: args is a JSON string, not an object.

Two calls instead of 26 tools cluttering the context.

Config

File Layout

Use the shared MCP files when you want one setup to work across hosts, and Pi-owned files when you need Pi-specific overrides or settings.

| File | Purpose | |------|---------| | ~/.config/mcp/mcp.json | User-global shared MCP config | | .mcp.json | Project-local shared MCP config | | <Pi agent dir>/mcp.json | Pi global override and compatibility imports (~/.pi/agent/mcp.json by default) | | .pi/mcp.json | Pi project override |

Pi-specific files are the write targets for imported or shared global servers when Pi needs to persist adapter-only settings such as directTools.

Server Options

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "some-mcp-server"],
      "lifecycle": "lazy",
      "idleTimeout": 10
    }
  }
}

| Field | Description | |-------|-------------| | command | Executable for stdio transport | | args | Command arguments | | env | Environment variables; supports ${VAR} and $env:VAR interpolation | | cwd | Working directory; supports ${VAR}, $env:VAR, and ~ expansion | | url | HTTP endpoint (StreamableHTTP with SSE fallback) | | headers | HTTP headers; supports ${VAR} and $env:VAR interpolation | | auth | "bearer" or "oauth" | | oauth.grantType | "authorization_code" (default) or "client_credentials" for non-interactive machine auth | | bearerToken / bearerTokenEnv | Token or env var name; bearerToken supports ${VAR} and $env:VAR interpolation | | lifecycle | "lazy" (default), "eager", or "keep-alive" | | idleTimeout | Minutes before idle disconnect (overrides global) | | exposeResources | Expose MCP resources as tools (default: true) | | directTools | true, string[], or false — register tools individually instead of through proxy | | excludeTools | string[] of tool names to hide (matches original names like get_screenshot and prefixed names like figma_get_screenshot) | | debug | Show server stderr (default: false) |

Lifecycle Modes

• lazy (default) — Don't connect at startup. Connect on first tool call. Disconnect after idle timeout. Cached metadata keeps search/list working without connections.
• eager — Connect at startup but don't auto-reconnect if the connection drops. No idle timeout by default (set idleTimeout explicitly to enable).
• keep-alive — Connect at startup. Auto-reconnect via health checks. No idle timeout. Use for servers you always need available.

Settings

{
  "settings": {
    "toolPrefix": "server",
    "idleTimeout": 10
  },
  "mcpServers": { }
}

| Setting | Description | |---------|-------------| | toolPrefix | "server" (default), "short" (strips -mcp suffix), or "none" | | idleTimeout | Global idle timeout in minutes (default: 10, 0 to disable) | | directTools | Global default for all servers (default: false). Per-server overrides this. | | disableProxyTool | Hide the mcp proxy tool once configured direct tools are fully available from cache. | | autoAuth | Auto-run OAuth on connect/tool calls when a server needs auth, then retry once (default: false). | | sampling | Allow MCP servers to sample through Pi models, honoring modelPreferences.hints before current/default fallback (default: true when UI approval is available). | | samplingAutoApprove | Skip sampling confirmation prompts. Required for sampling in non-UI sessions (default: false). |

Per-server idleTimeout overrides the global setting.

Direct Tools

By default, all MCP tools are accessed through the single mcp proxy tool. This keeps context small but means the LLM has to discover MCP tools via proxy search. If you want specific tools to show up directly in the agent's tool list — alongside read, bash, edit, etc. — add directTools to your config.

Per-server:

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"],
      "directTools": true
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "directTools": ["search_repositories", "get_file_contents"]
    },
    "huge-server": {
      "command": "npx",
      "args": ["-y", "mega-mcp@latest"]
    }
  }
}

| Value | Behavior | |-------|----------| | true | Register all tools from this server as individual Pi tools | | ["tool_a", "tool_b"] | Register only these tools (use original MCP names) | | Omitted or false | Proxy only (default) |

To set a global default for all servers:

{
  "settings": {
    "directTools": true
  },
  "mcpServers": {
    "huge-server": {
      "directTools": false
    }
  }
}

Per-server directTools overrides the global setting. The example above registers direct tools for every server except huge-server.

To exclude specific tools while still using directTools: true, add excludeTools on the server:

{
  "mcpServers": {
    "figma": {
      "url": "http://localhost:3845/mcp",
      "directTools": true,
      "excludeTools": ["get_figjam", "figma_get_code_connect_map"]
    }
  }
}

excludeTools filters direct tools, proxy search/list/describe, and the /mcp panel view.

Each direct tool costs ~150-300 tokens in the system prompt (name + description + schema). Good for targeted sets of 5-20 tools. For servers with 75+ tools, stick with the proxy or pick specific tools with a string[].

Direct tools register from the metadata cache in the Pi agent dir (~/.pi/agent/mcp-cache.json by default, or $PI_CODING_AGENT_DIR/mcp-cache.json when set), so no server connections are needed at startup. On the first session after adding directTools to a new server, the cache won't exist yet — tools fall back to proxy-only and the cache populates in the background. To force it: /mcp reconnect <server>.

When you change direct-tool toggles in /mcp or write new config through /mcp setup, the extension triggers Pi's normal reload flow automatically. That refreshes extensions, prompts, skills, and MCP tool registration in one shot, so newly configured direct tools can appear without a manual restart.

Interactive configuration: Run /mcp to open an interactive panel showing all servers with connection status, tools, and direct/proxy toggles. You can reconnect servers and toggle tools between direct and proxy from the same overlay. For OAuth, press Enter on a server that needs auth or ctrl+a on any OAuth server.

Guided first-run setup: Run /mcp setup to inspect detected shared MCP files, adopt compatibility imports from other hosts, open discovered config paths, preview exact before/after file diffs for writes, scaffold a minimal project .mcp.json, or quick-add RepoPrompt into a standard/shared MCP file.

Subagent integration: If you use the subagent extension, agents can request direct MCP tools in their frontmatter with mcp:server-name syntax. See the subagent README for details.

MCP UI Integration

MCP servers can ship interactive UIs via the MCP UI standard. When you call a tool that has a UI resource, the adapter opens it in a native macOS window via Glimpse if available, otherwise falls back to the browser.

How it works:

1. Agent calls a tool like launch_dashboard
2. The tool's metadata includes _meta.ui.resourceUri pointing to a UI resource
3. pi-mcp-adapter fetches the UI HTML and opens it in an iframe
4. The UI can call MCP tools and send messages back to the agent

Native rendering: On macOS, if Glimpse is installed (pi install npm:glimpseui), UIs open in a native WKWebView window instead of a browser tab. Set MCP_UI_VIEWER=browser to force the browser, or MCP_UI_VIEWER=glimpse to require native rendering.

Bidirectional communication: The UI talks back. When it sends a prompt or intent, the message is stored and triggerTurn() wakes the agent. The agent retrieves messages via mcp({ action: "ui-messages" }) and responds, enabling conversational UIs where the app and agent collaborate in real-time.

Session reuse: When the agent calls the same tool again while its UI is already open, the adapter pushes the new result to the existing window instead of replacing it. This enables live updates — the agent can refine a chart, add data, or respond to user input without losing the current view. Different tools still replace the session as before.

Message types from UI:

| Type | Purpose | |------|---------| | prompt | User message that triggers an agent response | | intent | Structured action with name + params | | notify | Fire-and-forget notification | | message | Generic message payload | | (custom) | Any other type forwarded as intent |

Retrieving UI messages:

mcp({ action: "ui-messages" })

Returns accumulated messages from UI sessions. Each message includes type, sessionId, serverName, toolName, and timestamp. Prompt messages include prompt, intent messages include intent and params.

Browser controls:

• Cmd/Ctrl+Enter — Complete and close
• Escape — Cancel and close
• Done/Cancel buttons — Same as keyboard shortcuts

Technical notes:

• Tool consent gates whether UIs can call MCP tools (never/once-per-server/always)
• Works with both stdio and HTTP MCP servers
• Uses a local 408KB AppBridge bundle (MCP SDK + Zod) for browser↔server communication

Local Example: Interactive Visualizer

A minimal MCP UI example at examples/interactive-visualizer demonstrating charts, bidirectional messaging, and streaming. From that directory:

npm install
npm run build
npm run install-local

Restart pi, then ask the agent to show a chart — it calls show_chart and opens the UI in Glimpse (macOS) or the browser. Use npm run uninstall-local to remove the MCP entry.

Import Existing Configs

Shared MCP files are loaded automatically. Use imports only for host-specific config formats that are not already covered by .mcp.json or ~/.config/mcp/mcp.json.

{
  "imports": ["cursor", "claude-code", "claude-desktop"],
  "mcpServers": { }
}

Supported compatibility imports: cursor, claude-code, claude-desktop, vscode, windsurf, codex

pi-mcp-adapter init detects these host-specific configs and adds missing imports to the Pi agent dir config for you.

Project Config

Prefer .mcp.json for project-local shared MCP config. Use .pi/mcp.json only when you need a Pi-specific project override. Project files override both user-global shared MCP config and Pi global overrides.

Usage

| Mode | Example | |------|---------| | Status | mcp({ }) | | List server | mcp({ server: "name" }) | | Search | mcp({ search: "screenshot navigate" }) | | Describe | mcp({ describe: "tool_name" }) | | Call | mcp({ tool: "...", args: '{"key": "value"}' }) | | Connect | mcp({ connect: "server-name" }) | | UI messages | mcp({ action: "ui-messages" }) |

MCP proxy and direct-tool results render compactly by default: long text shows the first three lines plus a Ctrl+O to expand hint, while the full result remains available when expanded and is still returned unchanged to the model.

Search includes both MCP tools and Pi tools (from extensions). Pi tools appear first with [pi tool] prefix. Space-separated words are OR'd.

Tool names are fuzzy-matched on hyphens and underscores — context7_resolve_library_id finds context7_resolve-library-id.

Commands

| Command | What it does | |---------|--------------| | /mcp | Interactive panel and first-run onboarding surface | | /mcp setup | Guided setup for imports, a minimal .mcp.json, RepoPrompt quick-add, and config-path inspection | | /mcp tools | List all tools | | /mcp reconnect | Reconnect all servers | | /mcp reconnect <server> | Connect or reconnect a single server | | /mcp-auth | Open an OAuth server picker in interactive UI sessions | | /mcp-auth <server> | OAuth setup for a specific server |

If settings.autoAuth is true, mcp({ connect: ... }), mcp({ tool: ... }), and direct tool calls automatically run OAuth when needed and retry once.

In interactive sessions, you can also authenticate from /mcp with ctrl+a or Enter on a server that needs auth. In non-interactive sessions, browser-based OAuth still requires /mcp-auth <server>. /mcp-auth without a server only opens a picker in the interactive UI.

How It Works

• One mcp tool in context (~200 tokens) instead of hundreds
• Servers are lazy by default — they connect on first tool call, not at startup
• Tool metadata is cached to disk so search/list/describe work without live connections
• Idle servers disconnect after 10 minutes (configurable), reconnect automatically on next use
• npx-based servers resolve to direct binary paths, skipping the ~143 MB npm parent process
• MCP server validates arguments, not the adapter
• Keep-alive servers get health checks and auto-reconnect
• Specific tools can be promoted from the proxy to first-class Pi tools via directTools config, so the LLM sees them directly instead of having to search

Limitations

• Cross-session server sharing not yet implemented (each Pi session runs its own server processes)
• Compact MCP result rendering summarizes text, but inline images are still controlled by Pi's image display settings and may render below the compact text summary.
• MCP sampling support is text-only; context inclusion, tools, stop sequences, audio, and image content are rejected with explicit errors.