Skip to content

Home

Jump to bottom
sylvain legland edited this page Mar 1, 2026 · 6 revisions

Software Factory — Your AI Software Engineering Team

Software Factory is a self-hosted platform that runs a team of 156 AI agents to design, build, test, and deploy software. You describe what you want; the platform coordinates the agents, tracks progress, manages costs, and delivers working code.

No cloud lock-in. No magic black box. Every decision is logged, every step is replayable.

🆕 What's New in v2.7.0

Feature What it does
Settings Hub Central control panel at /settings — mission concurrency (semaphore 1-10), budget cap per run, auto-resume watchdog, YOLO mode
Mission Cockpit Real-time pipeline view at /cockpit — watch every mission move through ideation → running → tests → deployed, with live semaphore gauge
LLM Cost Tracking Per-run llm_cost_usd field, budget alerts when a run approaches its cap, provider cost comparison dashboard
Auto-Resume Watchdog Paused missions restart automatically after a platform reboot — no manual intervention needed
YOLO Mode Fully autonomous runs: agents skip human validation checkpoints and self-approve. Enable per-run or globally in Settings
Semaphore Control Tune max concurrent missions live (1–10) from the Cockpit gauge or the Settings Hub — no restart required

🚀 Getting Started in 5 Minutes

Option A — Docker (recommended)

git clone https://github.com/macaron-software/software-factory.git
cd software-factory
make setup   # copies .env.example → .env, pulls images
make run     # starts all services
# Open http://localhost:8090

Option B — CLI

pip install sf-cli
sf init      # interactive setup wizard
sf start
# Open http://localhost:8090

Sign in with the admin account created during make setup (or sf init). You're ready to run your first mission.

🖥️ What You Can Do

🎯 Run a Mission

Ask the platform to build something. Open the Missions board (/missions), type a plain-English description, and submit. The orchestrator breaks the request into tasks, picks the right agents, runs them in order, and reports results.

  • Supports SAFe Epics, Features, Stories, and Tasks
  • Auto-assigns agents by skill (Frontend, Backend, QA, DevOps, Security, …)
  • Enable YOLO mode per-run for a fully hands-off execution

📡 Monitor Progress

  • Cockpit (/cockpit) — real-time pipeline: ideation → missions → running → tests → deployed. Live semaphore gauge. Per-mission controls: pause, resume, abort.
  • Metrics (/metrics) — token usage, cost per run (llm_cost_usd), agent performance, success rates, provider comparison.

⚙️ Configure the Platform

Settings Hub (/settings) gives you full control without touching config files:

Tab What you configure
Orchestrator Semaphore (1–10), budget cap, watchdog on/off, YOLO mode on/off
LLM Model routing, provider keys, per-agent model overrides
Integrations GitHub, GitLab, Jira, Slack, Datadog, and 25+ other tools
Notifications Slack/email/webhook alerts for mission events
Modules Toggle optional knowledge sources (19 available)

👥 Manage Your Team

  • RBAC (/rbac) — roles: Admin, Manager, Developer, Viewer. Assign per-workspace or globally.
  • Workspaces (/workspaces) — isolate projects, share agents, set per-workspace budgets.

🛒 Browse Agents & Workflows

  • Marketplace (/marketplace) — browse, install, and rate community agent packs and workflow templates.
  • Workflows (/workflows) — view and edit the 36 built-in workflows or build custom ones with the visual editor.
  • Tool Builder (/tool_builder) — create MCP-compatible tools for agents without writing boilerplate.

🔬 Evaluate Quality

  • Evals (/evals) — run standardized benchmarks against your agent pool. Track quality trends over time.
  • Mission Replay — step through any past mission turn-by-turn: every prompt, every response, every decision.
  • Backlog (/backlog) — triage failed or stalled missions, re-queue with adjusted parameters.

📸 Screenshots Gallery

Core Operations

Dashboard — at-a-glance platform status Dashboard

Missions Board — all missions in one kanban-style view Missions Board

Mission Cockpit — real-time pipeline with live semaphore gauge Mission Cockpit

Mission Replay — step-by-step audit of any past mission Mission Replay

Agent Management

Agents — browse all 156 agents, filter by domain and skill Agents

Marketplace — community agent packs and workflow templates Marketplace

Memory — agent long-term memory and knowledge base inspector Memory

Tool Builder — create MCP-compatible tools visually Tool Builder

Analytics & Quality

Metrics — token usage, LLM cost tracking, agent performance Metrics

Evals — benchmark runs and quality trends Evals

Backlog — triage stalled missions and re-queue Backlog

Platform Configuration

Settings Hub — mission concurrency, budget cap, YOLO mode, watchdog Settings Hub

Settings → Orchestrator — semaphore control and auto-resume watchdog Settings — Orchestrator

Settings → LLM — model routing and cost controls Settings — LLM

Settings → Integrations — 30+ tool connectors Settings — Integrations

Settings → Notifications — Slack, email, and webhook alerts Settings — Notifications

Settings → Modules — toggle optional knowledge sources Settings — Modules

RBAC — roles and permissions per workspace RBAC

Workspaces — project isolation and team management Workspaces

Workflows — built-in and custom workflow editor Workflows

OPS — infrastructure health and deployment controls OPS

Projects — project portfolio overview Projects

Integrations

3D World — live spatial view of agents and their interactions 3D World

📋 Key Concepts

Concept What it means in practice
Mission A unit of work you assign to the platform — from a one-liner ("add dark mode") to a full feature ("build the payment flow with Stripe"). The platform decomposes it, runs agents, and delivers a result.
Agent A specialized AI worker with a defined role (e.g. BackendDeveloper, SecurityAuditor, TechWriter). Each agent has its own system prompt, toolset, and memory.
Workflow A named sequence of agent calls — the recipe the platform follows when executing a mission. You can inspect, fork, and edit all 36 built-in workflows.
YOLO Mode Agents skip human validation checkpoints and auto-approve their own outputs. Great for fully automated CI pipelines. Use with caution in production.
Semaphore The maximum number of missions running simultaneously. Set to 1 for serial execution, up to 10 for maximum parallelism. Tunable live — no restart needed.
Watchdog A background process that monitors paused missions and resumes them automatically after a platform restart, so in-progress work is never lost.

🗺️ Navigation

Section Description
Architecture Platform architecture, components, data flow
Deployment Guide 3 environments: Azure, OVH, Local
API Reference REST API endpoints, authentication
Agents 156 agents across 9 domains
Workflows 36 built-in workflows
Patterns 15 orchestration patterns
Security Auth, adversarial validation, secrets
LLM Configuration Multi-model routing, Darwin LLM A/B, providers
Darwin Teams Evolutionary team selection + LLM Thompson Sampling
Evolution Genetic Algorithm + RL for workflow optimization
Integrations 30+ external tool connectors (all agent roles)
Guidelines Architecture guidelines: Wiki scraper + domain constraints

🌍 Translations

🇫🇷 Français · 🇪🇸 Español · 🇩🇪 Deutsch · 🇮🇹 Italiano · 🇧🇷 Português · 🇨🇳 中文 · 🇯🇵 日本語

📦 Quick Install

# Docker (recommended)
git clone https://github.com/macaron-software/software-factory.git
cd software-factory
make setup && make run
# → http://localhost:8090

# Or with the CLI
pip install sf-cli && sf init && sf start

🔒 For production deployments, see the Deployment Guide.

Repositories

Repo Purpose
macaron-software/software-factory Public GitHub repo — full platform, all agents and workflows
GitLab (internal) Enterprise skeleton — no mission data, CI/CD integration

License

AGPL-3.0 — See LICENSE

Clone this wiki locally