CLI Reference

Source: src/langgraph_kit/cli.py

langgraph-kit includes a CLI for scaffolding new agents from templates.

Usage

uv run python -m langgraph_kit.cli <command> [args]

Commands

new

Generate a new agent file from a template.

uv run python -m langgraph_kit.cli new <agent_id> [--output-dir <path>]

Argument	Required	Default	Description
agent_id	Yes	—	ID for the new agent (e.g., my-agent)
--output-dir	No	.	Directory to write the generated file

Output: Creates <agent_id>.py with a full agent template including:

• Prompt section definitions (core identity, etc.)
• Worker definitions (researcher, implementer, verifier)
• Tool registration using builder utilities
• Middleware stack construction
• Backend factory setup
• Command dispatcher configuration
• build_graph() function following the standard contract

Example:

uv run python -m langgraph_kit.cli new code-reviewer --output-dir backend/src/app/agents/graphs/

list

Show available agent templates.

uv run python -m langgraph_kit.cli list

Currently shows one template: default.

Generated Agent Structure

The generated agent file includes commented sections you can customize:

# Prompt sections — customize the agent's identity and instructions
_CORE_SECTIONS = [
    PromptSection(id="core_identity", content="...", stability=SectionStability.STABLE),
    ...
]

# Worker definitions — customize sub-agent roles
WORKER_DEFINITIONS = [
    {"name": "researcher", "system_prompt": "..."},
    {"name": "implementer", "system_prompt": "..."},
    {"name": "verifier", "system_prompt": "..."},
]

# Build function — the entry point
def build_graph(checkpointer, store):
    ...

After generating, register the agent in graphs/__init__.py to include it in register_all().

openapi

Dump the FastAPI agent router's OpenAPI specification to stdout or a file. Useful for generating typed clients without spinning up a live server.

uv run python -m langgraph_kit.cli openapi [--output spec.json] [--indent 2]

Argument	Required	Default	Description
--output	No	stdout	Write the spec to this path instead of stdout
--indent	No	2	JSON indent for the dumped spec (0 for compact)

What gets exported: the full router from langgraph_kit.contrib.fastapi.create_agent_router() mounted on a temporary FastAPI app. The spec covers /agents, /invoke, /stream, and any other routes the router declares, with all Pydantic request/response models referenced under components.schemas.

Generating a typed client

The dumped spec feeds into any OpenAPI client generator. Two recipes — pick the one that matches your stack.

Python client via openapi-python-client:

# 1. Dump the spec.
uv run python -m langgraph_kit.cli openapi --output spec.json

# 2. Generate a typed client package next to your app.
uvx --from openapi-python-client openapi-python-client generate --path spec.json

The generator writes a package whose Client class wraps every endpoint with typed attrs-style models. Drop it into your app and call Client(base_url=...).agents_get_invoke(...).

TypeScript / JS client via openapi-generator-cli:

uv run python -m langgraph_kit.cli openapi --output spec.json
npx @openapitools/openapi-generator-cli generate \
    -i spec.json \
    -g typescript-fetch \
    -o ./gen/langgraph-kit-client

Notes

• SSE doesn't model cleanly in OpenAPI. The /stream endpoint declares text/event-stream as its response content type; concrete SSE event payload schemas are referenced from components.schemas so generated clients can still type the events even though the transport itself is opaque.
• No live server required. The CLI mounts the router on a temporary FastAPI() and calls app.openapi(); nothing binds to a port. Safe to run in CI to keep a generated SDK up to date.

shell

Interactive REPL for a registered agent. See langgraph-kit shell --help for the full flag list.

Slash command	Effect
/exit, /quit, /q	End the session (or Ctrl-D / Ctrl-C).
/info	Print the active agent id, thread id, user id, and module path without invoking the agent.