Flet MCP Server

MCP server that makes the installed Flet source code the source of truth for AI agents.

Flet changes faster than model training data. Controls get renamed, deprecated and removed (ft.ElevatedButton → ft.Button, and Button(text=...) is now Button(content=...)), properties change types, icon names get hallucinated. The only reliable reference is the flet source code actually installed in your environment — this server ships with flet as a dependency, gives your AI tools to read it directly, and — uniquely — can run and verify AI-written Flet code against that installed flet before you ever see it.

Upgrading from v0.2.x? The plain command below works again — no --with mcp<2 --with typing_extensions flags needed anymore. v0.2.x broke because the mcp SDK 2.0 removed mcp.server.fastmcp; v1.0.0 is built on the new MCPServer API and pins mcp>=2.

uvx flet-mcp-server

Verify your Flet code (the killer feature)

verify_flet_code(code) checks AI-generated Flet code against the installed flet in two passes and reports every problem with line numbers and hints:

• Static: unknown controls, properties that don't exist anymore, enum typos, deprecated classes, undefined event handlers.
• Dynamic: the code runs in a sandboxed subprocess (app launchers are neutralized — nothing opens, nothing starts) while flet's own validators fire on every constructed control, catching errors that normally only appear at runtime — like Slider(min=10, max=5).

{
  "status": "errors",
  "flet_version": "0.86.5",
  "controls_verified": 6,
  "diagnostics": [
    {"severity": "warning", "code": "deprecated", "line": 4,
     "message": "'ElevatedButton' is deprecated in flet 0.86.5.",
     "hint": "Did you mean 'FilledButton'?"},
    {"severity": "error", "code": "bad-kwarg", "line": 4,
     "message": "'ElevatedButton' has no property 'text' in flet 0.86.5."},
    {"severity": "error", "code": "runtime", "line": null,
     "message": "ValueError: Slider: min (10) must be less than or equal to max (5)"}
  ]
}

The server's instructions tell AI clients to verify before delivering and fix-and-reverify until the report is passed — plus three ready-made prompts (verify_flet_code_prompt, migrate_flet_prompt, build_flet_ui_prompt) that encode the full workflow.

What your AI gets

Verify (new in v1.0.0) — confidence that AI-written code is actually valid:

• verify_flet_code(code) — static + sandboxed dynamic verification against the installed flet, structured diagnostics with line numbers and hints.
• Prompts: verify_flet_code_prompt, migrate_flet_prompt, build_flet_ui_prompt.
• Resources: flet://version and flet-source://<module> (browsable installed source, path-traversal safe).

Source of truth — reads the flet package installed alongside the server, so every answer matches the exact version your app runs:

• inspect_flet_control(control_name) — every property with type, default and origin class (inherited fields marked), on_* events, deprecation warnings, the class hierarchy, and the full class source with per-property docstrings.
• get_flet_version() — which flet version the tools are reading, and from where.
• search_flet_source(query) — grep the installed flet sources; definitions rank first.
• read_flet_source(module, symbol?) — numbered source of any module, or a single class/function extracted by AST.
• list_flet_api() — the true flet.__all__ of the installed version, grouped by category (Material / Cupertino / Core controls, Services, Components & hooks).
• search_flet_icons(query, icon_set) — valid ft.Icons.* / ft.CupertinoIcons.* names, so the AI never invents one again.
• search_flet_colors(query) — valid ft.Colors.* constants, including shades like AMBER_500.

Docs & examples — official guides and runnable apps:

• search_flet_docs(query) — smart search (direct > keyword aliases > fuzzy).
• get_flet_doc(doc_path, offset?, max_lines?) — Markdown of a doc page, paged so long pages don't flood the conversation.
• list_flet_controls() — every control that has a docs page.
• search_flet_examples(query) — official example apps (counter, todo, 7guis, routing, games…) straight from the flet repo, always current.
• get_flet_example(example_id) — full source of an example project.

Ecosystem — packages beyond core flet:

• list_official_packages() — official extensions from the monorepo (flet-audio, flet-video, …).
• search_flet_ecosystem(query) — verified community packages on GitHub/PyPI.
• get_package_details(package_name) — PyPI version, classification, install command.

Tools reference (complete list)

All 16 tools, three prompts and two resources exposed by the server:

Prompts: verify_flet_code_prompt(code), migrate_flet_prompt(project_summary), build_flet_ui_prompt(spec) — guided workflows that encode the verify-first loop.

Resources: flet://version (static) and flet-source://{module} (template — installed source, path-traversal safe).

Matching your project's Flet version

By default the source tools read the latest flet bundled with the server. Two ways to verify against the exact version your project uses:

1. Pin the version at launch (matches uvx-style workflows):

uvx --with flet==0.86.1 flet-mcp-server

2. Local Mode — point the tools at your project's virtualenv, so they read its flet (and only its flet) no matter what the server bundles:

{
  "mcpServers": {
    "flet-mcp-server": {
      "command": "uvx",
      "args": ["flet-mcp-server"],
      "env": { "FLET_MCP_VENV": "/absolute/path/to/your/project/.venv" }
    }
  }
}

Every tool response then reports e.g. [flet 0.86.5 — FLET_MCP_VENV=…/project/.venv], and inspect_flet_control shows deprecations relative to that version.

Client Configuration Examples

VSCode

Add this to your .vscode/mcp.json:

{
  "servers": {
    "flet-mcp-server": {
      "command": "uvx",
      "args": ["flet-mcp-server"]
    }
  }
}

Antigravity

Add this to your mcp_config.json:

{
  "mcpServers": {
    "flet-mcp-server": {
      "command": "uvx",
      "args": ["flet-mcp-server"]
    }
  }
}

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "flet-mcp-server": {
      "command": "uvx",
      "args": ["flet-mcp-server"]
    }
  }
}

Cursor / Windsurf

In your IDE's MCP settings, add a new server:

• Name: Flet MCP
• Type: Command
• Command: uvx flet-mcp-server

Zed

Add this to your settings.json file inside the context_servers object:

{
  "flet": {
    "command": "uvx",
    "args": ["flet-mcp-server"],
    "env": {}
  }
}

OpenCode

Add this to your ~/.config/opencode/opencode.json or project-level .opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "flet-mcp": {
      "type": "local",
      "command": ["uvx", "flet-mcp-server"],
      "enabled": true
    }
  }
}

Configuration

Docker / remote deployments: run with FLET_MCP_TRANSPORT=streamable-http FLET_MCP_HOST=0.0.0.0 — a GET /health liveness endpoint is included. OpenTelemetry traces are emitted automatically by the MCP SDK; configure any standard OTEL_* environment variables to export them.

Development

git clone https://github.com/Nwokike/flet-mcp-server.git
cd flet-mcp-server
uv sync

uv run pytest                              # unit tests (75)
uv run ruff check src/ tests/ scripts/     # lint
uv run python scripts/smoke_stdio.py       # full MCP handshake from a fresh uvx env

The smoke script is the regression test for the v0.2.0 outage: it installs the server into a clean ephemeral environment (exactly what end users get) and verifies the handshake plus one call per tool group.

See docs/ARCHITECTURE.md for the full architecture and CHANGELOG.md for release history.

License

MIT