I'm Diego, and I built architect to solve a problem that gets bigger every month: AI coding agents are incredibly powerful, but they still need someone sitting in front of them. Claude Code, Cursor or Copilot, they're brilliant for interactive pair-programming. But what about the 16 hours a day when nobody's at the keyboard? That's the gap architect fills. It's a CLI tool designed from day one to run without supervision, in your CI/CD pipeline, in a cron job, overnight while you sleep. The agent works, tests verify the result, and you wake up to a PR that actually passes. Here's what makes it different: 🔁 Ralph Loop --> autonomous iteration until tests pass. This is the killer feature. architect runs your task, then runs your tests and linters. If they fail, it starts a fresh iteration, clean context, no accumulated garbage from previous attempts. Just the original spec, the accumulated diff, and the errors. It keeps going until everything passes. Not until the LLM thinks it's done, until your pytest, ruff, or tsc says it's done. architect loop "Implement the payments module" \ --check "pytest tests/ -q" \ --check "ruff check src/" \ --max-iterations 25 --max-cost 3.00 ⚡ Parallel runs with git worktrees. Launch 3 models on the same task and compare real results in your codebase. Or fan-out different tasks in parallel. Each worker gets its own isolated git worktree, no copies, no Docker, no conflicts. 🛡️ Deterministic guardrails --> Protected files the AI can't touch. Blocked commands. Code rules enforced on every write. Quality gates that must pass before the agent can declare "done." All declarative in YAML. The LLM can't skip them, ever. 📋 Declarative pipelines --> Multi-step workflows in YAML: Plan → Build → Test → Review → Fix. Each step is a separate agent with its own model, context, and checks. Variables flow between steps. Checkpoints let you resume from any point. 🪝 Hooks on 10 lifecycle events. Pre/post tool use, pre/post LLM call, session start/end. Auto-format after every edit. Block dangerous commands. Notify Slack when the agent finishes. Your logic, injected into every step. 🔀 100+ models via LiteLLM --> OpenAI, Anthropic, Google, DeepSeek, Mistral, Azure Foundry, Ollama, any OpenAI-compatible endpoint. Switch models with one line. Zero vendor lock-in. 🏗️ CI/CD-native --> Exit codes (0=success, 1=failed, 2=partial). JSON output. Budget and timeout as hard limits. Markdown reports for PR comments. No interactive prompts. architect is a step in your GitHub Actions, not a chat window. 🧠 Learns with use --> Procedural memory auto-detects your corrections and saves them. Next time, it already knows your conventions. Plus .architect.md for project rules and Vercel-compatible skills. 📊 Full observability --> Human-readable logs with icons for real-time monitoring. Execution reports in JSON/Markdown. Native OpenTelemetry traces for Grafana, Datadog, Jaeger. What makes architect different from Claude Code, Cursor, or Aider? Those are interactive tools, brilliant when you're at the keyboard. Architect is for when you're not. It's the missing piece between "I have an AI that generates code" and "I have an AI that delivers verified code at 3am without anyone watching." What's next? The core is solid and everything above works today. But the roadmap is ambitious. Next up: architect as an MCP Server, so Claude Code or any MCP-compatible agent can invoke architect as a tool, launch a pipeline, run a loop, get a report, all without leaving their workflow. Bidirectional: architect orchestrates agents, and agents orchestrate architect. Also building a backend abstraction layer so you can swap between LiteLLM and Claude Agent SDK without changing a single guardrail, pipeline, or hook. Same control layer, different engine underneath. On the governance side: OWASP-aligned guardrail presets out of the box, compliance reports mapped to EU AI Act categories, and cost intelligence that analyzes your spend across models, tasks, and time, all from the execution reports architect already generates. The goal is clear: make architect the trust layer that teams need before they let AI agents touch production code unsupervised. Not the smartest agent in the room, the most reliable one. Architect is fully open source (MIT), installs with pip install architect-ai-cli, and works on Linux, macOS, and Windows. Python 3.12+. I'd love your feedback, especially on which CI/CD integrations or pipeline recipes to build next. Star the repo if autonomous coding agents in your pipeline is something you've been waiting for!

Architect

The AI coding agent that works while you sleep.

The AI coding agent that works while you sleep.