Langflow

Components are organized as Core components grouped by purpose and Bundles grouped by service provider, covering language models, embeddings, vector stores, document loaders, data sources and tools.

An API Request component calls arbitrary endpoints, custom Python components extend the catalog arbitrarily, and Tool Mode on any component exposes it to an Agent component as a callable tool. As an MCP client Langflow connects to external MCP servers and uses their tools.

Provider bundles documented include Apify actors and search providers, and vector database and model provider bundles span the major commercial and local options.

Sourcedocs.langflow.org components overview, bundles, MCP client and API request pagesread 2026-08-31

Flows are node graphs of typed components with documented loop and branching behavior, parsing and type conversion between incompatible ports, and dynamic ports that open as templates reference new variables.

Components can be grouped into a single reusable component and saved to the component menu, and freezing a component pins it and all upstream components so their output is reused rather than recomputed. The Agents section documents building and configuring agents and using components and flows as agent tools, which allows one agent to invoke another flow. Individual components can be run in isolation to test dependencies.

Sourcedocs.langflow.org agents, components overview and flows pagesread 2026-08-31

Retrieval is assembled from documented component classes: document loaders, text splitters such as the Recursive Character Text Splitter with configurable chunk size and overlap, embedding model components, and vector store components supplied as provider bundles so the customer chooses and owns the store.

Retrieval components feed context into prompt templates through typed ports, with a Type Convert component bridging incompatible types. Uploaded files are handled by a documented file management system that stores names, paths, sizes and storage providers. Chat memory can itself be backed by an external vector store.

Sourcedocs.langflow.org components overview, vector store bundles, memory management and file management pagesread 2026-08-31

Human-in-the-Loop is documented as pausing a flow, creating a checkpoint and waiting for a human decision, after which the run resumes from the checkpoint along the selected branch without re-executing completed steps.

Two mechanisms are documented: the Human Input component, which pauses the flow where it is placed and creates one output branch per configured user action such as Approve or Reject, and per-tool approval on the Agent component, where enabling Requires approval on an individual tool pauses the run only when the agent attempts to call that tool. The capability is documented as a Langflow 1.11 release feature.

Sourcedocs.langflow.org human-in-the-loop, human-input component and agent tools pagesread 2026-08-31

The vendor documents API key and JWT authentication, external authentication through an upstream OIDC proxy or corporate SSO gateway with JWKS validation and automatic user provisioning, and a pluggable authorization layer with viewer, developer and admin roles, shares and an authorization audit log.

The docs state plainly that open-source Langflow registers a pass-through authorization service and ships no enforcement plugin, so per-resource RBAC is not enforced without a third-party plugin. No security attestation applies to the self-hosted open-source project, and access enforcement is left to the operator.

Sourcedocs.langflow.org authorization and external-authentication pagesread 2026-09-29

Each component exposes an Inspect control showing that component's output and logs for the most recent run, and a Last Run indicator confirms execution, so a builder can see what every node in the graph produced. Monitor endpoints expose stored messages programmatically at GET /v1/monitor/messages filterable by session ID. The Playground shows message logs per session with full history, and log level and file settings are configurable through environment variables. No Langflow documentation page describes tracing integrations with LangSmith or Langfuse.

Sourcedocs.langflow.org components overview, API reference and playground pagesread 2026-08-31

Memory bases store an agent's long-term chat history in a vector store for semantic retrieval of past conversations across sessions, filling themselves from each attached flow's runs through auto-capture and a configurable ingestion threshold, one memory base per flow and several per flow allowed, on Chroma, Chroma Cloud, OpenSearch or pgvector. Chat conversations are also stored by session ID, with the Agent component's chat memory on by default and a Message History component with store and retrieve modes, and message logs are editable and deletable.

Sourcedocs.langflow.org memory-bases and session-id pagesread 2026-09-29

Langflow is an MIT-licensed open-source Python application that the customer self-hosts, with a documented Deploy section covering containerized deployment and serving flows over a network, installation as a Python package, and Langflow Desktop as a downloadable application for macOS and Windows that bundles dependencies.

LFX is separately documented for building Langflow as a dependency within another application. Storage is a database the operator controls, and LANGFLOW_USE_NOOP_DATABASE disables persistence entirely for testing. Chat memory can be directed to external databases the customer already runs.

Sourcedocs.langflow.org deployment overview, LFX overview, memory management pages and langflow.org desktopread 2026-08-31

The vendor states that Langflow includes several pre-built templates that are ready to use or customize, and documentation procedures routinely begin by creating a flow from a named template such as Basic Prompting or the Memory Chatbot starter example. New workspaces open with a Starter Project folder by default. Grouped components can be saved into the Core components menu as reusable custom components, and flows are importable and exportable as JSON so complete assets can be shared and adopted.

Sourcedocs.langflow.org get started, memory, web-search and import and export pagesread 2026-08-31

A Webhook component gives each flow its own POST /v1/webhook endpoint, authenticated by API key by default, so an external application's event starts a run without a person; the docs index lists running flows via the API or webhooks. Channels include the REST API, an embeddable langflow-chat web component, the Playground and flows published as MCP tools. No built-in scheduler or cron trigger is documented.

Sourcedocs.langflow.org llms.txt index and sitemap (webhook and component-webhook pages)read 2026-09-29

A Language Model component selects provider and model per node, with the documentation routinely instructing users to add an OpenAI key or select a different provider and model.

Model providers are supplied as Bundles grouped by provider covering the major commercial options including OpenAI, Anthropic, Google, Cohere and Hugging Face, alongside local models through Ollama.

The component can emit either a model response or a Language Model output that powers a downstream component's reasoning, so one flow can drive several models. Custom Python components allow any additional provider to be added.

Sourcedocs.langflow.org components overview, model components and web-search pagesread 2026-08-31

Langflow exposes flows as MCP servers consumable by external MCP clients and also acts as an MCP client connecting to external MCP servers, both documented in a dedicated Model Context Protocol section. A documented REST API runs flows, lists components at /api/v1/all, triggers runs and returns monitor data, authenticated by API key. Flows import and export as JSON, custom components are written in Python and shareable, an embeddable langflow-chat web component integrates flows into a site with host URL and API key props, and LFX is documented for building Langflow as a dependency.

Sourcedocs.langflow.org MCP, API reference, import and export, custom components and publish pagesread 2026-08-31

Documented testing and debugging affordances are the Playground for interactive flow testing with custom session IDs, running a single component in isolation via build_vertex without executing upstream dependencies, Inspect for per-component output and logs, Freeze to prevent a component and its upstream from re-running while preserving last output, editable and deletable message logs whose modification changes subsequent chatbot behavior, and automatic backup flows before applying component updates that may contain breaking changes. No evaluation, scoring, regression or benchmarking capability is documented.

Sourcedocs.langflow.org playground, components overview and memory pagesread 2026-08-31

No first-party browser, desktop or computer-use component is documented. The nearest capability is the Web Search component, which consolidates web, news and RSS modes and which the vendor itself describes as using web scraping subject to rate limits, alongside an API Request component for arbitrary HTTP calls. Browser-adjacent capability reaches Langflow only through third-party routes: an Apify bundle that calls a hosted RAG Web Browser actor, and community-authored custom components wrapping external browser-automation libraries.

Sourcedocs.langflow.org web-search and components referenceread 2026-08-31