alfasteine

The standalone CLI for ALFASTEIN — a browser-native agent team builder with one-command deployment.

Install

npm install -g alfasteine

Quickstart

1. Build your team in the browser

Open the ALFASTEIN studio at /agent-studio, configure your team (model per vertex, tools per role), and click Deploy.

2. Paste the install command anywhere

The Deploy modal generates a one-liner:

npx alfasteine init --config=eyJ2ZXJzaW9uIjoxLC...

This decodes your team configuration and creates a new project directory with package.json, README.md, .env.example, and alfastein.config.json.

3. Run the agent

cd <team-name>
cp .env.example .env  # fill in API keys for your providers
alfasteine run --prompt "Your task here"

Note: As of v0.1.0, alfasteine run requires running from inside the ALFASTEIN monorepo. Standalone runtime is on the roadmap — see Phase 23.

Commands

• alfasteine init --config=<base64> — decode a TeamConfig and scaffold a project directory
• alfasteine init --config=<base64> --cwd=<dir> — scaffold into a specific directory
• alfasteine init <name> — create a single agent from a template (non-team mode)
• alfasteine init <name> --interactive — launch the 7-step setup wizard
• alfasteine run — boot the agent against a team config (requires monorepo)
• alfasteine run --vertex <id> — run a specific vertex from the team
• alfasteine run --prompt "..." — one-shot prompt

Architecture

• 11 LLM providers with auto-fallback: OpenAI, Anthropic, Google, OpenRouter, NovitaAI, Ollama, z.ai/GLM, Kimi/Moonshot, Arcee, MiniMax, xAI
• 10 builtin tools: read_file, write_file, web_search, shell, ssh_shell, daytona_shell, screenshot, visual_diff, mcp_call, spawn_subagent
• Auto-triggered skills (DeepSeek visual-primitives doctrine fires on UI/layout/render keywords)
• Multi-turn ReAct loop with iteration budget + grace call + AbortSignal interrupts
• Token quantization: Anthropic prompt cache stratification, per-role output budgets, optional Haiku-tier tool result compression
• SSE token streaming for OpenAI- and Anthropic-compatible providers
• Recursive subagent spawning (depth ≤ 3) with full parent-context inheritance
• Sandbox backends: local shell, remote ssh_shell (allowlisted hosts + commands), daytona_shell (pre-provisioned Daytona workspaces)

License

MIT. Copyright © 2026 Yassine Drani.