MCP Server Setup

This page documents the per-client install matrix: the exact configuration snippet each MCP-compatible client needs to spawn ./bin/dowel mcp:serve (or dart run :dispatcher mcp:serve as a fallback) against your project. The dowel MCP server is stdio-only in V1 (no remote HTTP endpoint); every client below launches the CLI locally with the project directory as the working directory.

Without dowel MCP#

LLM agents working against a running Flutter app are blind without dowel MCP:

• No access to the live VM Service: cannot read Magic.find<X>().rxState or any controller state.
• No way to drive the UI: gestures, snapshots, screenshots, modal waits all need the Tapir extension surface.
• No HTTP / log / exception inspection: Crumbs's runtime feed stays invisible.
• No connected REPL: every state inspection becomes a print() round-trip + hot-reload.

With dowel MCP#

The dowel MCP server bridges the agent to the running app over the VM Service:

• 10 substrate command tools (dowel_start, dowel_stop, dowel_status, dowel_logs, dowel_restart, dowel_reload, dowel_hot_restart, dowel_doctor, dowel_list, dowel_tinker).
• Plugin tools surface automatically once the consumer's bin/dispatcher.dart wrapper registers the provider; see each plugin's MCP tool reference for the current catalogue (tapir, crumbs).
• Tool visibility filters via .dowel/mcp.json + env vars + CLI flags (Cargo-style replace for allow, union for deny).
• Zero release-build impact: every dowel integration gates on kDebugMode at the consumer's main.dart.

Add use dowel to your prompt to nudge the agent toward the MCP tool surface:

Show me the current MonitorController state. use dowel

Installation#

Prerequisites#

The MCP server ships with dowel. Add the package + scaffold the consumer wrapper first so ./bin/dowel and bin/dispatcher.dart exist:

dart pub add dowel
dart run dowel install

install writes bin/dispatcher.dart + bin/dowel + the codegen barrels (one-time per project; idempotent on re-run). After that, mcp:install writes the .mcp.json entry the clients below consume.

Canonical entry shape#

The canonical .mcp.json entry shape (written by ./bin/dowel mcp:install or dart run dowel mcp:install):

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Every client below uses the canonical shape above, though the command + args payload adjusts based on platform and dowel availability. The cwd field must point at the project root (the directory that contains pubspec.yaml).

Cursor#

Go to: Settings, Cursor Settings, MCP, Add new global MCP server.

Paste into ~/.cursor/mcp.json:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

For project-scoped use, write the same snippet to .cursor/mcp.json in the project root. The spawned process is equivalent to running ./bin/dowel mcp:serve from that directory.

Claude Code#

Run this command from the project root to add the dowel MCP server:

claude mcp add dowel -- ./bin/dowel mcp:serve

For project-scoped configuration:

claude mcp add --scope project dowel -- ./bin/dowel mcp:serve

(When bin/dowel is absent or on Windows, use dart run :dispatcher mcp:serve instead.)

You can also write .mcp.json directly with the canonical entry shape from the Installation section above.

Note: Claude Code does NOT auto-reconnect on .mcp.json edits. Run /mcp reconnect dowel inside Claude Code after every edit to load the new entry.

Claude Desktop#

On Windows and macOS, install the official Claude Desktop application.

Open File, Settings, Developer, MCP Servers, Edit Config to open claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the dowel MCP server configuration. Replace the cwd value with the absolute path to your Flutter project root, since Claude Desktop has no project context:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "/absolute/path/to/your/flutter/project"
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Important: Fully quit Claude Desktop via File, Exit (closing the window just minimises it). Restart it for the new entry to take effect.

VS Code (GitHub Copilot)#

Create or edit .vscode/mcp.json in your project:

{
  "servers": {
    "dowel": {
      "type": "stdio",
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

For user-level configuration, add the same entry to your VS Code settings under chat.mcp.servers.

Windsurf#

Add to your Windsurf MCP configuration:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Reload the Cascade panel after editing the config; Windsurf spawns ./bin/dowel mcp:serve on next request.

JetBrains IDEs (Junie / AI Assistant)#

For Junie: Open Junie, click the three dots in the top right corner, then Settings, MCP Settings:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

For AI Assistant: Go to Settings, Tools, AI Assistant, MCP and add the same entry via the "as JSON" option.

Cline / Roo-Code#

Add to your Cline MCP server configuration (typically ~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json on macOS, equivalent paths on other OSes):

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Roo-Code reads the same mcpServers shape; point its config file at the snippet above.

OpenCode#

Add to your OpenCode configuration file (~/.config/opencode/opencode.json or per-project opencode.json):

{
  "mcp": {
    "servers": {
      "dowel": {
        "type": "local",
        "command": ["./bin/dowel", "mcp:serve"],
        "cwd": "."
      }
    }
  }
}

OpenCode takes the executable and its arguments from the one command array. When bin/dowel is absent or on Windows, use ["dart", "run", ":dispatcher", "mcp:serve"] instead; that is the entry mcp:install writes for the other clients.

The local transport spawns the process directly; cwd pins the project root.

Gemini CLI#

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Restart gemini after editing; the CLI re-spawns ./bin/dowel mcp:serve on the next session.

Antigravity#

Antigravity is Google's AI-powered IDE built on VS Code.

1. Open the Agent side panel (Cmd/Ctrl + L or View, Open View..., Agent).
2. Click Additional options (...) in the upper right of the Agent panel.
3. Select MCP Servers.
4. Click Manage MCP Servers.
5. Click View raw config in the upper right of the Manage MCPs editor view.
6. Add this configuration:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Tip: Install the Dart and Flutter extensions for the best development experience.

Firebase Studio#

Firebase Studio is an agentic cloud-based development environment by Google.

1. In your Firebase Studio project, create .idx/mcp.json if it does not exist.
2. Add the dowel MCP configuration:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

3. Rebuild your workspace:
   • Open the Command Palette (Shift + Ctrl + P).
   • Enter Firebase Studio: Rebuild Environment.

Other Clients#

For any other MCP-compatible client that supports stdio transport, spawn the server with:

./bin/dowel mcp:serve

When bin/dowel is absent or on Windows, spawn with dart run :dispatcher mcp:serve instead.

The working directory must contain a pubspec.yaml that depends on dowel (directly or via a path / git dep). Clients that require a JSON entry shape accept the canonical form:

{
  "mcpServers": {
    "dowel": {
      "command": "./bin/dowel",
      "args": ["mcp:serve"],
      "cwd": "."
    }
  }
}

When bin/dowel is absent or on Windows, mcp:install writes {"command": "dart", "args": ["run", ":dispatcher", "mcp:serve"]} instead.

Remote HTTP transport is NOT supported in V1; the dowel MCP server is stdio-only because it bridges to the local VM Service URI recorded in this project's session by ./bin/dowel start (or dart run :dispatcher start when bin/dowel is absent).

Server Identity#

The dowel MCP server advertises itself as serverInfo: {"name": "dowel_mcp", "version": "<package version>"} in the initialize response. The name field is fixed regardless of the installed plugin set. Clients such as Claude Code, Cursor, Windsurf, and others derive tool prefixes (like mcp__dowel__dowel_start) from the .mcp.json key (typically dowel), not from serverInfo.name. The MCP specification treats serverInfo.name as a display hint, not a stable identifier for tool prefix derivation.

The version field in serverInfo tracks the package version at release time, manually synced as part of the release cut process. See the MCP specification at https://modelcontextprotocol.io/specification/2025-03-26/basic/lifecycle for the full lifecycle semantics. This addresses issue #9 NIT 7, which called out the version drift problem.

Add a Rule#

To make the agent reach for dowel MCP without an explicit use dowel prompt, add a rule to your AI client:

Cursor: Settings, Rules.

Claude Code: Add to CLAUDE.md in the project root.

Windsurf: Memories panel, add a workspace rule.

JetBrains AI Assistant: Settings, Tools, AI Assistant, Custom Instructions.

Example rule:

Always use dowel MCP when working with this Flutter project. Inspect running app state via dowel_tinker, drive the UI via tapir_* tools, and read HTTP / log / exception feeds via crumbs_*. Prefer dowel_status before assuming the app is running.

Automated Install#

The dowel CLI ships a built-in command that writes the canonical .mcp.json entry for you:

dart run dowel mcp:install

The command is idempotent: pre-existing mcpServers keys are preserved, and running it twice replaces the dowel key in-place without creating duplicates. By default it writes to .mcp.json in the current working directory; override the path via --path. The command always writes the mcpServers shape, which fits clients that read that key (Claude Code, Cursor, Windsurf); VS Code reads servers, so write its file by hand as shown above:

dart run dowel mcp:install --path ./config/mcp.json

To remove the entry later:

dart run dowel mcp:uninstall

Both commands work without a running Flutter app; they only edit the JSON file.

Available Tools#

The dowel MCP server surfaces 10 substrate tools plus plugin-contributed tools from any registered sibling plugin (tapir, crumbs, or custom plugins). See tool-reference.md for the per-tool input schema, output shape, and example invocations of the substrate tools, plus links to each plugin's MCP tool reference for the current plugin tool catalogue.

Tool visibility is filtered in three layers (file .dowel/mcp.json lowest priority, env vars middle, CLI flags highest priority; deny always wins over allow). To temporarily hide a tool surface without uninstalling the server, edit .dowel/mcp.json:

{
  "packages": {
    "allow": null,
    "deny": []
  },
  "tools": {
    "allow": null,
    "deny": []
  }
}

After editing .dowel/mcp.json or env vars, run /mcp reconnect dowel inside Claude Code (or the equivalent reconnect action in your client); the server does NOT auto-reload in V1.

Related Documentation#

• MCP Overview: the substrate and plugin tool layers.
• MCP Tool Reference: every substrate tool and its input.
• mcp:serve: the server command and its filter flags.