Usage¶

Most people use the CLI. Some teams prefer the web UI. The API is for wiring it into your own tooling. All three call the same generator.

CLI — the short version¶

agent-generator "<your sentence>" -f <framework> [-o <path>]

That's the only shape you need to remember.

Recipes¶

Generate a CrewAI research team¶

agent-generator "Research team with a researcher and a writer" \
  -f crewai -o research_team/

You get a full CrewAI project — crew.py, agents.yaml, tasks.yaml, tool stubs, tests, pyproject.toml, README.

Export a WatsonX Orchestrate agent¶

agent-generator "Customer support assistant that triages tickets" \
  -f watsonx_orchestrate -o support.yaml

orchestrate agents import -f support.yaml

Build a LangGraph data pipeline on OpenAI¶

agent-generator "Pipeline that extracts, transforms, loads data" \
  -f langgraph -p openai --model gpt-4o -o pipeline.py

Wrap any Python output as an MCP server¶

agent-generator "Research assistant" -f crewai --mcp -o assistant.py
python assistant.py          # FastAPI on :8080, POST /invoke

Estimate cost before generating¶

agent-generator "Multi-step content workflow" -f crewai_flow --show-cost

Try without credentials¶

agent-generator "Anything" -f react --dry-run

All the CLI flags¶

Flag	Default	What it does
`-f, --framework`	(required)	`crewai`, `crewai_flow`, `langgraph`, `react`, `watsonx_orchestrate`
`-p, --provider`	`watsonx`	`watsonx` or `openai`
`--model`	provider default	Override the model.
`--mcp / --no-mcp`	off	Add a FastAPI `/invoke` wrapper to Python output.
`-o, --output`	stdout	Write to a file or directory.
`--dry-run`	off	Skip the LLM call (fake spec, real renderer).
`--show-cost`	off	Print token + cost estimate before generating.
`--temperature`	0.7	Sampling temperature for the planning call.
`--max-tokens`	4096	Cap on the planning response.

Web UI¶

uvicorn agent_generator.wsgi:app --host 0.0.0.0 --port 8000 --reload

Open http://localhost:8000 and you get:

A text area to describe the agent team.
A framework picker (or Auto).
An artifact-mode picker (code, YAML, or both).
A tool catalogue (six prebuilt templates).
A file tree + syntax-highlighted preview after generation.
A Download ZIP button.

The full platform (make start) adds auth, projects, runs, marketplace, and audit on top of the same screens.

REST API¶

Path	Method	What it does
`/api/plan`	POST	Run the planner; get a `ProjectSpec` JSON.
`/api/build`	POST	Render files from a `ProjectSpec`.
`/api/generate`	POST	Plan + build in one call.
`/download/{id}`	GET	Download a generated bundle as ZIP.
`/health`	GET	Liveness probe.

curl -X POST http://localhost:8000/api/generate \
  -H 'content-type: application/json' \
  -d '{"prompt": "Research team", "framework": "crewai", "provider": "watsonx"}'

Docker¶

docker build -t agent-generator .
docker run -e WATSONX_API_KEY=... -e WATSONX_PROJECT_ID=... \
           -p 8000:8000 agent-generator

CLI inside the container:

docker run --rm -e WATSONX_API_KEY=... \
  agent-generator agent-generator "Say hi" -f react --dry-run

Next: Pick a framework · Architecture