fastapi-architect-mcp

An MCP server that gives Claude Code IDE-level intelligence for FastAPI projects — semantic code navigation, safe renaming, route inspection, dependency trees, Pydantic model analysis, and a typed knowledge graph of the whole application.

Instead of Claude reading files blindly, it calls structured tools backed by Jedi (Python language server) and Python's AST.

Tools

Knowledge graph

Knowledge graph

The graph is built statically from the AST — no code is executed and no LLM is involved.

Nodes: App, Router, Route, Handler, Dependency, Middleware, Function, Schema, ORMModel, Class, Table, Template, Module

Edges:

App/Router ──INCLUDES──▶ Router          (prefixes → full route paths)
Route ──HANDLED_BY──▶ Handler
Route/Router/function ──DEPENDS_ON──▶ Dependency
function ──ACCEPTS / RETURNS──▶ Schema/ORMModel
function ──CALLS──▶ function             function/class ──USES──▶ class
function ──QUERIES──▶ Table              (raw SQL strings: SELECT, INSERT, UPDATE, DELETE, CREATE)
function ──RENDERS──▶ Template           App ──MIDDLEWARE──▶ Middleware
ORMModel ──MAPS_TO / REFERENCES──▶ Table ORMModel ──RELATES_TO──▶ ORMModel
Schema ──INHERITS──▶ Schema              Schema ──MIRRORS──▶ ORMModel (name + field overlap, with confidence)

It covers both structured projects (APIRouter, Depends, SQLAlchemy/SQLModel) and flat ones (a single main.py, raw psycopg2 queries, Jinja2 templates, manual auth checks).

Cache: per-file extraction results are stored in <project>/.fastapi-architect/graph.json (git-ignored automatically). Only new or modified files are re-parsed, so graph tools stay fast after the first build.

Installation

pip install fastapi-architect-mcp

Configuration

Option 1 — Global (available in all projects)

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "fastapi-architect": {
      "command": "fastapi-architect-mcp"
    }
  }
}

Option 2 — Per project (recommended for best results)

Create .mcp.json at your project root:

{
  "mcpServers": {
    "fastapi-architect": {
      "command": "/path/to/your/project/venv/bin/fastapi-architect-mcp"
    }
  }
}

Using your project's own venv gives Jedi access to all your project's dependencies, which enables full cross-file find_references and rename_symbol support.

Then restart Claude Code.

Usage examples

Once connected, you can ask Claude Code things like:

• "List all routes in this project"
• "What are the dependencies of the `get_users` handler?"
• "Find all references to `get_db` across the project"
• "Rename `get_db` to `get_session` everywhere"
• "Inspect the `UserCreate` model"
• "Are there any routes missing a response_model?"
• "Show me the full dependency graph for the users router"
• "Are there any mismatches between my `User` ORM model and `UserPublic` schema?"
• "Give me an overview of this FastAPI project"
• "If I change the `User` model, which endpoints are affected?"
• "Which write endpoints have no authentication?"
• "How does `POST /chat` reach the `chat_logs` table?"
• "Export the knowledge graph as HTML"

Supported patterns

Dependency injection

• def handler(x=Depends(func)) — standard FastAPI
• def handler(*, x=Depends(func)) — keyword-only args
• @router.get("/", dependencies=[Depends(func)]) — decorator-level deps
• SessionDep = Annotated[Session, Depends(get_db)] — Annotated aliases (modern FastAPI)

Models

• BaseModel — Pydantic
• SQLModel — SQLModel
• BaseSettings — Pydantic settings
• RootModel — Pydantic v2
• Inherited models (e.g. class UserCreate(UserBase))

Validators

• @validator — Pydantic v1
• @field_validator — Pydantic v2

Known limitations

• `find_references` and `rename_symbol` work cross-file only when the MCP is configured to use the project's own venv (Option 2 above). With a global install, they only operate on the file where the symbol is defined.
• `get_dependencies` and the knowledge graph support native FastAPI Depends() / Security() only. Other DI frameworks (Dishka, dependency-injector, etc.) are not supported.
• Auth detection in audit_graph is inferred from names (dependencies, auth headers, called functions). Middleware-based auth is not detected; pass auth_dependencies to declare custom guards.
• Static analysis limits: routes registered dynamically, dependencies passed through variables, calls through self or injected objects, and SQL built at runtime are only partially captured.
• `export_graph_html` loads vis-network from a CDN, so viewing the page requires internet access.
• Runtime dependencies (middleware, lifespan events, startup hooks) are not visible to static analysis and won't appear in dependency trees.
• `get_completions` is most useful at attribute access positions (e.g. after a .). On empty lines it returns all Python builtins.

Requirements

• Python 3.11+
• Claude Code (VS Code extension or CLI)

Author

Abdessamad Touzani

• GitHub: @AbdessamadTzn
• LinkedIn: abdessamadtouzani