LangStitch Documentation

Build graph applications in Python, Java and Go

Use the Python SDK to build LangGraph applications with decorators, YAML and a CLI. Use the Spring AI compiler for Java, or LangTailor for native Go project export. Share graph topology and state while writing each target's native code explicitly. The support record describes the portable core and the features that still differ across languages.

Get started Quickstart API reference

Published release set

• Python SDK 0.3.2: install from PyPI, including real MCP tool clients for stdio, SSE and streamable HTTP.
• Java compiler 0.2.2: Maven Central, with native functions, HTTP tools and Spring AI project generation.
• LangTailor 0.4.0: Windows and macOS downloads, VS Code Marketplace and Open VSX, with native Python, Java and Go export.
• Shared spec 2.2.0: schemas and package archive with lossless native bodies and runtime settings. Distributed through GitHub Releases.

Publication and generated-application CI were verified separately from production acceptance. Read the release notes for the evidence and remaining limits.

Start with Python

pip install langstitch-sdk            # core (PyYAML only)
pip install "langstitch-sdk[all]"     # + FastAPI server + LangGraph + LangChain + httpx

langstitch new my-agent
cd my-agent
pip install -e ".[server,graph,llm,http]"
python -m app          # bootstrap + print registered components
langstitch run         # start the API server

What you get

• Decorators

  ---

  Register nodes, graphs, skills, tools, agents, guardrails, policies, personas, and MCP primitives with one decorator each — bare or parameterized.

  Decorators

• YAML configuration

  ---

  Drive the app from application.yaml / env.yaml, read values with a JSON-path API, and bind typed sections to dataclasses.

  Configuration

• Graphs & nodes

  ---

  Wire nodes and subgraphs with GraphBuilder and compile to a real LangGraph StateGraph.

  Graphs & nodes

• Hierarchical context

  ---

  Every LLM / sub-agent call runs in an isolated child scope — only the final output merges back, so parents stay small.

  Context & registries

• External services

  ---

  Declare downstream HTTP services with auth + header propagation and call them through a typed client.

  External services

• MCP clients and metadata

  ---

  Call real MCP tools over stdio, SSE and streamable HTTP. Register server tools, resources and prompts as metadata for your server integration.

  MCP

Design principles

• Light core. Importing langstitch never imports FastAPI, LangGraph, LangChain, or httpx. Heavy libraries are opt-in extras; each helper imports its dependency lazily and raises a clear install hint if it is missing.
• Convention over configuration. A scaffolded project gives every concept its own package, and a single import app wires the whole application by registering decorators at import time.
• Declarative first. Behavior lives in YAML + decorated functions, so the same description can power the SDK, the LangTailor IDE, and Python codegen.

LangStitch vs LangTailor

LangStitch is the SDK on this site. LangTailor is the desktop IDE that designs the same graphs visually and exports Python, native Java and Go projects. Review the supported features and limits before selecting a target.

Product engineering

Read the native runtime release notes, production readiness record, and gap audit. The SDK overview links package installation instructions and language-specific export guides.