Open Source — MIT License

Build what you imagine.

An AI-assisted development environment for building functional, documented, and tested software. Runs in VS Code with Claude Code.

Download for macOS View on GitHub

VS Code Extension MCP Native Requires Claude Code

The Context Drift Nightmare

AI coding tools work great on small tasks. On larger projects, a pattern emerges.

Initial Magic AI understands everything

Isolation Success Individual pieces work

Integration Hell Nothing fits together

Context Collapse AI forgets prior decisions

Vision Drift Result doesn't match intent

This isn't a capability problem. LLMs suffer 30%+ accuracy drops in long context, and on repository-level tasks, no model exceeds 30% success. The AI doesn't have the right context at the right time.

Every session starts from scratch. Without a shared knowledge base, the AI re-discovers patterns, re-reads files, and re-solves problems. Your original vision gets lost.

Why not just use Claude Code?

Claude Code is powerful. Crucible makes it structured.

More throughput

Specialized agents operating in parallel — requirements, design, implementation, testing, security, DevOps — each working simultaneously on different tasks.

Better quality

Structured process and documentation at every stage. Quality gates enforce review before handoff. Every decision is traceable from intent to code.

Compounding efficiency

A project knowledge graph that grows as you build. Prior requirements inform future design. Past decisions guide new implementation. The system gets smarter with every task.

How It Works

You direct the vision. Crucible handles the structure. Every task makes the system smarter.

You work with specialized agents to create and resolve tasks — producing requirements, designs, code, and tests for a fully functional output. Each agent has a full chat interface and access to specialized tools.

1 Agents — specialized AI personas 2 Tasks — create, assign, and resolve work items 3 Chat — full conversation interface per agent

As you work, the system creates structured documents modeled on how the best software companies operate — requirements, designs, architecture decisions, code, and tests. It then learns from this knowledge base and becomes smarter the more you use it.

The Virtuous Cycle — Guided Process, Detailed Artifacts, Knowledge Graph, Intelligent Context, LLM Efficiency, Enhanced Tools

Task Management

Break work into trackable tasks in VS Code. See progress, manage dependencies, track blockers.

Knowledge Graph

Artifacts indexed and searchable. Context carries forward between sessions — the AI never starts from scratch.

Persona Workflows

Specialized roles — design, implementation, testing — each with appropriate tool access and workflow guidance.

Quality Gates

Validation before handoff. Signoff enforcement. Structured review that catches issues before they compound.

Semantic Search

Embeddings + full-text search across all project artifacts. Prior work informs future tasks.

Vision Preservation

Your original intent traces from requirements to design to code to verification. The AI knows why code exists, not just what it does.

Who It's For

Good fit

Developers building multi-file projects in VS Code
Codebases too large for a single AI conversation
Work that benefits from structured handoffs and discoverable decisions
Anyone who's hit the wall where AI tools lose context

Probably not

Quick scripts or single-file tasks — Claude Code alone is fine
No VS Code or terminal experience
Want fully autonomous AI — Crucible is a partner, not an autopilot

Crucible Alpha vs Pro

Alpha is available now. Pro is in active development.

Crucible Alpha

Available Now

Built on Claude Code. Single LLM, full file system access. Best-effort documentation — does a good job but not locked in by design.

Requires Claude Code
Specialized AI personas
Knowledge graph & semantic search
Quality gates & traceability
VS Code extension
Open source (MIT)

Crucible Pro

In Development

LLM-agnostic. Adaptive model selection for speed, cost, quality. Predictable file structures, locked-in quality gates, enforced process.

Multi-LLM support — any provider
Adaptive model selection
State-machine orchestration (30+ sub-graphs)
Agent swarming & auto-scaling pools
Self-improving system capabilities
Persistent sessions per persona

Get Started

Alpha is a v1 built fast — functional, real productivity gains, rough in places. It works. Read the full getting started guide →

1. Download and unzip

Download Crucible for macOS — unzip and open the folder

2. Open in VS Code

code Crucible-1.0.0

3. Start working

# Claude Code auto-connects via .mcp.json — tools available immediately.

Requires: VS Code Claude Code Node.js 20+

Alternative: install from source

git clone https://github.com/ember-agentic/crucible && cd crucible && npm run install:all