agent-generator

Type a sentence. Ship a multi-agent project — as code, a container, a desktop app, or an Android build.

---

Two surfaces, one repo

Surface	What you get	Use when…
CLI	pip install agent-generator → one-shot project generator (this README, ↓ further down)	You want a single Python file or YAML.
Platform	FastAPI backend + Vite SPA + Tauri desktop + Capacitor Android, all from one TypeScript codebase.	You want the full app, the wizard, the Marketplace, signed installers, or a Helm install.

Run the full platform in three commands

make install      # CLI + backend (FastAPI) + frontend (Vite SPA)
make test         # CLI tests + backend pytest + frontend type-check & build
make start        # backend on :8000 and SPA on :5173 — Ctrl-C stops both

Then open http://localhost:5173. Ports are configurable (make start BACKEND_PORT=8080 FRONTEND_PORT=4173).

Ship desktop / Android binaries

make build              # nginx SPA bundle + backend Docker image
                        # + .dmg / .msi / .AppImage / .deb / .rpm for THIS host
make build-android      # Android debug APK (needs Android SDK + JDK 17)

Signed CI builds (Authenticode, Apple notarization, GPG, Play Store) live in .github/workflows/{desktop,mobile,release-app}.yml.

For the production Kubernetes deployment, see deploy/helm/.

---

What the CLI does

You type what you want. You get a complete project.

agent-generator "Build a research team with a researcher and a writer" -f crewai

That's it. You now have working Python code with agents, tasks, and tools.

---

Try it online

No install needed. Open the demo and start building:

https://huggingface.co/spaces/ruslanmv/agent-generator

---

Install

pip install agent-generator

That gives you the CLI and all core features. Python 3.10+ required.

Optional extras:

pip install "agent-generator[openai]"     # OpenAI provider
pip install "agent-generator[all]"        # All providers + frameworks

---

Quick start

Step 1 -- Set credentials (pick one)

# Option A: WatsonX (default)
export WATSONX_API_KEY=your-key
export WATSONX_PROJECT_ID=your-project-id

# Option B: OpenAI
export OPENAI_API_KEY=sk-your-key
export AGENTGEN_PROVIDER=openai

Step 2 -- Generate

# CrewAI team
agent-generator "Research team that finds papers and writes summaries" -f crewai -o team.py

# LangGraph pipeline
agent-generator "ETL pipeline: extract, transform, load" -f langgraph -o pipeline.py

# WatsonX YAML agent
agent-generator "Customer support assistant" -f watsonx_orchestrate -o support.yaml

# Just test it (no credentials needed)
agent-generator "Hello world agent" -f crewai --dry-run

Step 3 -- Run

python team.py

---

Supported frameworks

Framework	What you get
CrewAI	crew.py + agents.yaml + tasks.yaml + tools + tests
LangGraph	graph.py with StateGraph and typed state
WatsonX Orchestrate	agent.yaml ready for orchestrate agents import
CrewAI Flow	Flow class with @start / @listen
ReAct	Reasoning loop with tool registry

---

Built-in tools

Your agents can use these out of the box:

Tool	What it does
web_search	Search the web
pdf_reader	Read PDF files
http_client	Call APIs
sql_query	Query databases
file_writer	Save files
vector_search	Semantic search

---

Web UI

A visual wizard that guides you step by step:

uvicorn agent_generator.wsgi:app --port 8000

Open http://localhost:8000 -- describe your agents, pick a framework, download a ZIP.

Works with local LLMs (Ollama), remote LLMs (OllaBridge), or cloud APIs (OpenAI).

---

CLI options

agent-generator "your description" [options]

Flag	What it does
-f crewai	Pick framework: crewai, langgraph, watsonx_orchestrate, crewai_flow, react
-o file.py	Save to file (otherwise prints to screen)
--dry-run	Generate without calling any LLM
--mcp	Add a FastAPI server wrapper
-p openai	Use OpenAI instead of WatsonX
--show-cost	Show estimated cost before generating

---

Docker

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

---

Contributing

git clone https://github.com/ruslanmv/agent-generator.git
cd agent-generator

# Full stack (CLI + backend + frontend) in one shot:
make install
make test
make start            # backend :8000 + SPA :5173

# Or just the CLI bits, like the original workflow:
pip install -e ".[dev,all]"
pytest                # CLI tests
ruff check src/       # lint
mkdocs serve          # docs

Useful make targets:

Target	What it does
make install	Editable CLI install plus make app-install (backend + SPA)
make test	CLI pytest plus backend pytest + frontend tsc + Vite build
make start	Backend + frontend dev servers concurrently (Ctrl-C stops both)
make stop	Stop any backend / frontend dev servers started by make start
make build	Frontend bundle + backend Docker image + desktop installer (host)
make build-android	Android debug APK
make help	Full list (CLI + platform targets, ~30 of them)

The full architecture (compatibility matrix, frontend shells, backend service, signing pipeline, Helm chart) is documented in docs/platform.md. The original master plan is in docs/complete-solution-plan.md.

---

License

Apache 2.0 -- PRs and issues welcome.