Quickstart

1. Write an agent

A LangStage agent is a LangGraph CompiledGraph exported from a Python file or module. Nothing LangStage-specific is required.

No model, no keyA real model (deepagents + Anthropic)

A plain LangGraph graph. It needs only langgraph, which every surface already installs, so it runs as-is:

my_agent.py

from langchain_core.messages import AIMessage
from langgraph.graph import END, START, MessagesState, StateGraph

def respond(state):
    last = state["messages"][-1].content
    return {"messages": [AIMessage(content=f"You said: {last}")]}

g = StateGraph(MessagesState)
g.add_node("respond", respond)
g.add_edge(START, "respond")
g.add_edge("respond", END)
graph = g.compile()

langstage-cli init writes this same file, plus a langstage.toml that points at it.

Needs extra packages and an API key

pip install deepagents langchain-anthropic and set ANTHROPIC_API_KEY. The surfaces don't install deepagents for you (except the web app's langstage[deepagents] extra).

my_agent.py

from deepagents import create_deep_agent

graph = create_deep_agent(model="anthropic:claude-sonnet-4-6")

You don't need to compile in a checkpointer: every surface attaches one when the graph has none (the web app's is durable SQLite; see Web).

2. Point any stage at it with one spec string

The agent spec is path/to/file.py:attr or module:attr. Every stage understands the same form:

langstage run --agent my_agent.py:graph          # web
langstage-cli --agent my_agent.py:graph          # terminal  (-a for short)
langstage-jupyter --agent my_agent.py:graph      # JupyterLab launcher (-a too)
langstage-agui --agent my_agent.py:graph         # AG-UI endpoint

Prefer not to pass it each time? Set it once in the environment or in langstage.toml (see Configuration), and every stage picks it up:

export LANGSTAGE_AGENT_SPEC="my_agent.py:graph"

langstage.toml

[agent]
spec = "my_agent.py:graph"   # relative to this file, so it works from any subdirectory

3. Check it before you chat

Three questions, in order:

langstage-cli --show-config                        # does the config resolve my agent?
langstage-cli --verify                             # does it load and run one turn?
langstage-cli "What can you do?"                   # what does it say?

The same trio exists on each surface (--show-config, then langstage check --live / --verify / --selfcheck, then a one-shot message). See Installation for the full list.

The resolved-config table shows each value, where it came from, and the env var and TOML key that set it:

Resolved config  (value  [source]):

  agent_spec       = /home/you/project/my_agent.py:graph [toml (langstage.toml)]   (env: LANGSTAGE_AGENT_SPEC (legacy DEEPAGENT_AGENT_SPEC), toml: agent.spec)
  workspace_root   = .                          [default]   (env: LANGSTAGE_WORKSPACE_ROOT (legacy DEEPAGENT_WORKSPACE_ROOT), toml: workspace.root)
  graph_name       = graph                      [default]   (toml: agent.graph_name)
  verbose          = False                      [default]   (toml: ui.verbose)
  (not used by this surface, so not shown: host, port, debug, title, stream_mode, async_mode)

  TOML read from: /home/you/project/langstage.toml
  ...

4. Try a stage with no API key

Every surface has a keyless demo, good for a first look, a screenshot or a test:

langstage run --demo
langstage-cli --demo "ping"
langstage-jupyter --demo
langstage-vscode-sidecar --demo --message "ping"
langstage-hermes demo
langstage-agui --demo

To see tool calls, reasoning and an approval prompt without a key, use the tool demo: langstage-cli -a langstage_core.demo.tools:graph "use a tool" (or "think about it", or "ask me").

On the web stage, the empty screen offers a few starter-prompt chips. Click one to begin without typing.

Now pick the stage you want: Web · Terminal · JupyterLab · VS Code · Reference agent