AGENTS.md

This file provides guidance to AI coding agents when working with code in this repository.

Project Overview

Langflow is a visual workflow builder for AI-powered agents. It has a Python/FastAPI backend, React/TypeScript frontend, and a lightweight executor CLI (lfx).

Prerequisites

• Python: 3.10-3.14
• uv: >=0.4 (Python package manager)
• Node.js: >=20.19.0 (v22.12 LTS recommended)
• npm: v10.9+
• make: For build coordination

Common Commands

Development Setup

make init              # Install all dependencies + pre-commit hooks
make run_cli           # Build and run Langflow (http://localhost:7860)
make run_clic          # Clean build and run (use when frontend issues occur)

Development Mode (Hot Reload)

make backend           # FastAPI on port 7860 (terminal 1)
make frontend          # Vite dev server on port 3000 (terminal 2)

For component development, enable dynamic loading:

LFX_DEV=1 make backend                    # Load all components dynamically
LFX_DEV=mistral,openai make backend       # Load only specific modules

Code Quality

make format_backend    # Format Python (ruff) - run FIRST before lint
make format_frontend   # Format TypeScript (biome)
make format            # Both
make lint              # mypy type checking

Testing

make unit_tests                    # Backend unit tests (pytest, parallel)
make unit_tests async=false        # Sequential tests
uv run pytest path/to/test.py      # Single test file
uv run pytest path/to/test.py::test_name  # Single test

make test_frontend                 # Jest unit tests
make tests_frontend                # Playwright e2e tests

Database Migrations

make alembic-revision message="Description"  # Create migration
make alembic-upgrade                         # Apply migrations
make alembic-downgrade                       # Rollback one version

Architecture

Monorepo Structure

src/
├── backend/
│   ├── base/langflow/     # Core backend package (langflow-base)
│   │   ├── api/           # FastAPI routes (v1/, v2/)
│   │   ├── components/    # Built-in Langflow components
│   │   ├── services/      # Service layer (auth, database, cache, etc.)
│   │   ├── graph/         # Flow graph execution engine
│   │   └── custom/        # Custom component framework
│   └── tests/             # Backend tests
├── frontend/              # React/TypeScript UI
│   └── src/
│       ├── components/    # UI components
│       ├── stores/        # Zustand state management
│       └── icons/         # Component icons
├── langflow-core/         # Usable provider-free Langflow distribution
├── bundles/               # Curated provider integrations
└── lfx/                   # Lightweight executor and shared primitives

Key Packages

• langflow: Full end-user package; depends on langflow-core and curated provider bundles
• langflow-core: Service-complete, provider-bundle-free distribution; owns the langflow CLI
• langflow-base: Modular application platform (API, services, graph engine); extras add service integrations
• lfx: Shared execution primitives and standalone CLI (lfx serve, lfx run)

The public dependency direction is langflow → langflow-core → langflow-base → lfx. Provider packages under src/bundles/ are added only by the full langflow distribution.

Service Layer

Backend services in src/backend/base/langflow/services/:

• auth/ - Authentication
• authorization/ - Authorization (RBAC) plugin layer — see below
• database/ - SQLAlchemy models and migrations
• cache/ - Caching layer
• storage/ - File storage
• tracing/ - Observability integrations

Authorization (RBAC)

Authorization is a pluggable layer separate from authentication:

• OSS ships the interface (BaseAuthorizationService in lfx) + a pass-through implementation (LangflowAuthorizationService) + the authz_* and casbin_rule DB schema + route guards.
• Implementations register via the lfx.services entry point authorization_service in lfx.toml (same pattern as the SSO auth_service). A registered plugin reads the authz_* admin tables and writes compiled rules to casbin_rule.

Default is off: LANGFLOW_AUTHZ_ENABLED=false. When enabled with only the OSS stub registered, every check returns allow — the stub is a no-op so routes stay wired and audit rows still flow. Real allow/deny requires a registered authorization plugin.

Route guards live in langflow.services.authorization.guards (the legacy langflow.services.authorization.utils path re-exports them for backward compatibility):

• ensure_flow_permission(user, FlowAction.*, flow_id=..., flow_user_id=..., workspace_id=..., folder_id=...) — single-flow CRUD + execute
• ensure_deployment_permission(user, DeploymentAction.*, deployment_id=..., deployment_user_id=..., workspace_id=..., project_id=...)
• ensure_project_permission(user, ProjectAction.*, project_id=..., project_user_id=..., workspace_id=...)
• ensure_knowledge_base_permission(user, KnowledgeBaseAction.*, kb_name=..., kb_user_id=...)
• ensure_variable_permission(user, VariableAction.*, variable_id=..., variable_user_id=...)
• ensure_file_permission(user, FileAction.*, file_id=..., file_user_id=...)
• ensure_share_permission(user, ShareAction.*, share_id=..., share_user_id=...)
• filter_visible_resources(user, resource_type=..., candidates=..., act=...) — list-endpoint filter; safe no-op in OSS

The enforcement request shape is (subject, domain, object, action):

• subject = user:{uuid}
• domain = project:{uuid} → workspace:{uuid} → * (resolved by _resolve_flow_domain; the more specific domain wins so project-scoped grants match directly while workspace-scoped grants still flow down via plugin-side role inheritance)
• object = flow:{uuid} / deployment:{uuid} / project:{uuid} / flow:* / etc.
• action = read / write / create / delete / execute / deploy

Share-aware fetch (Phase 3): route fetch helpers (_read_flow, get_flow_by_id_or_endpoint_name, get_deployment, project reads in projects.py, v2 file fetcher, variable PATCH/DELETE in variable.py) branch on BaseAuthorizationService.supports_cross_user_fetch(). The OSS pass-through reports False so the existing owner-scoped queries are preserved — enabling LANGFLOW_AUTHZ_ENABLED=true without a registered plugin cannot widen visibility. Plugins set SUPPORTS_CROSS_USER_FETCH=True so resources load by id alone and ensure_*_permission decides access; route handlers can convert a plugin-deny HTTPException(403) to HTTPException(404) via langflow.services.authorization.fetch.deny_to_404 to preserve UUID privacy.

Share CRUD API (Phase 3): /api/v1/authz/shares provides POST / GET / PATCH / DELETE on authz_share rows. The handler enforces an OSS floor (resource owner or superuser may administer shares for that resource) so the OSS pass-through cannot let a non-owner mint share rows. Each write fires BaseAuthorizationService.invalidate_user / invalidate_all so a registered enforcer can drop cached policy. Audit rows are written via audit_decision with share:create / share:update / share:delete actions.

Audit query API (Phase 4): GET /api/v1/authz/audit (superuser-only) exposes a paginated, filterable view of authz_audit_log. Supports user_id, resource_type, resource_id, action, result, since, until filters; page size capped at 200.

Default role catalog (Phase 4): the consolidated foundations migration 7c8d9e0f1a2b_authz_foundations seeds the three built-in is_system=True roles (viewer / developer / admin) with "{resource}:{action}" permission slugs. OSS does not interpret these — they exist so a registered plugin's policy sync has a stable bootstrap source.

Component Development

Components live in src/backend/base/langflow/components/. To add a new component:

1. Create component class inheriting from Component
2. Define display_name, description, icon, inputs, outputs
3. Add to __init__.py (alphabetical order)
4. Run with LFX_DEV=1 make backend for hot reload

IMPORTANT: Changing a component's class name is a breaking change and should never be done. The class name serves as an identifier used to match components in saved flows and to flag them for updates in the UI. Renaming it will break existing flows that use that component.

Component Structure

from langflow.custom import Component
from langflow.io import MessageTextInput, Output

class MyComponent(Component):
    display_name = "My Component"
    description = "What it does"
    icon = "component-icon"  # Lucide icon name or custom

    inputs = [
        MessageTextInput(name="input_value", display_name="Input"),
    ]
    outputs = [
        Output(display_name="Output", name="output", method="process"),
    ]

    def process(self) -> Message:
        # Component logic
        return Message(text=self.input_value)

Component Testing

Tests go in src/backend/tests/unit/components/. Use base classes:

• ComponentTestBaseWithClient - Components needing API access
• ComponentTestBaseWithoutClient - Pure logic components

Required fixtures: component_class, default_kwargs, file_names_mapping

Frontend Development

• React 19 + TypeScript + Vite
• Zustand for state management
• @xyflow/react for graph visualization
• Tailwind CSS for styling

Custom Icons

1. Create SVG component in src/frontend/src/icons/YourIcon/
2. Export with forwardRef and isDark prop support
3. Add to lazyIconImports.ts
4. Set icon = "YourIcon" in Python component

Testing Notes

• @pytest.mark.api_key_required - Tests requiring external API keys
• @pytest.mark.no_blockbuster - Skip blockbuster plugin
• Database tests may fail in batch but pass individually
• Pre-commit hooks require uv run git commit
• Always use uv run when running Python commands
• When running tests inside a sub-package (e.g. langflow-base, lfx), sync that package's dev group first: uv sync --group dev --package langflow-base. The default uv sync only resolves the top-level workspace and may leave dev-only test deps (e.g. fakeredis) uninstalled.

Graph Testing Pattern

Proper Graph tests follow this pattern:

1. Build graph with connected components
2. Connect them via .set() calls
3. Call async_start and iterate over the results
4. Validate the results

Testing Best Practices

• Avoid mocking in tests when possible
• Prefer real integrations for more reliable tests

Version Management

make patch v=1.5.0  # Update version across all packages

This updates: pyproject.toml, src/backend/base/pyproject.toml, src/frontend/package.json

Pre-commit Workflow

Pre-commit hooks run ruff and biome automatically on git commit, so manual formatting is not required. To avoid an extra commit cycle when you have many changes:

1. Run make format_backend once before staging - fixes most ruff issues up front.
2. Run uv run git commit (the uv run ensures pre-commit finds the right Python).
3. If you touched backend code, run make unit_tests locally for faster feedback than CI.

Pull Request Guidelines

• Follow semantic commit conventions
• Reference any issues fixed (e.g., Fixes #1234)
• Ensure all tests pass before submitting

Documentation

Documentation uses Docusaurus and lives in docs/:

cd docs
yarn install
yarn start        # Dev server on port 3000 (prompts for 3001 if 3000 is in use)