Install
"One command in the terminal. It handled Node.js, the VS Code extension, and set up a workspace with all the personas. Took about a minute."
The installer runs through 9 phases — pre-flight checks verify macOS, Node.js, disk space, and Git. Once the install path is configured, it handles the rest: VS Code extension, workspace creation, and a health check.
First look
"I opened VS Code and there was Crucible in the sidebar — a list of AI personas, each with a specialty. Planning, Requirements, Architecture, Implementation, Testing. Like a team waiting for a brief."
The sidebar shows all available personas — Planning, Requirements, Architecture, Design, Implementation, Testing, and more. The task board is empty and ready. The terminal waits for a conversation.
Start with Planning
"I clicked the Planning persona. A Claude Code terminal opened — connected to Crucible's MCP server, with access to all the workflow tools. Ready for a conversation."
Each persona gets its own Claude Code session connected to Crucible's MCP server, with access to task creation, project structuring, and delegation tools.
Describe what you want
"I described the restaurant, the problems they're having, and what I wanted to build. I didn't tell it what technology to use or what artifacts to create — just the problem and what I needed. Then I asked it to figure out the scope before planning anything."
The prompt describes the problem and outcome, not the technology. "I don't know what I don't know yet. Before we plan anything, help me understand the full scope."
It asks before it plans
"Instead of jumping straight to task creation, it asked me questions. Real ones — about no-show policy, payment processing, whether we need deposits. It was doing discovery, not just generating."
The planner does real discovery — asking about no-show prevention, payment policies, and operational details. Concrete options with trade-offs for each question, not just yes/no. Understand first, plan second.
A structured plan emerges
"After the discovery conversation, it created a project structure: 4 planning tracks running in parallel, 9 requirements clusters handed off to the Requirements persona, which will produce 44 user stories covering the full scope. The task board filled up."
The task board fills up. 9 requirement clusters handed off to Requirements, 4 parallel planning tracks, 44 user stories covering the full scope. Not a flat list — a structured decomposition with phases, dependencies, and delegation.
Check the pulse
"I asked for a status report. It pulled live data — a pipeline table showing which personas had work, what was signed off, what was blocked. Not a summary — the actual state."
Live pipeline data from the MCP server — pipeline table with task counts per persona, completion and signoff progress, and items needing attention flagged in red.
Guardrails that work
"This is where it got interesting. The Requirements persona tried to hand off directly to Design, skipping Architecture. Crucible blocked it. Not a warning — it literally couldn't proceed. The agent figured out the right path on its own."
The agent tries an invalid handoff — Crucible blocks it. The pipeline flow is enforced: requirements → architecture → design → implementation → testing. The agent self-corrects and routes to Architecture. No human intervention needed.
The system learns as it goes
"Before creating anything new, the Architecture persona searched the knowledge base. It found existing requirements, checked for prior decisions, and built on what was already there. Every task adds to the knowledge graph — and every search uses it."
MCP tool calls: get_requirement retrieves prior decisions, search queries for "authentication OAuth JWT", list_architecture_specs checks what exists. The agent searches before creating — prior decisions inform future work.
Real architecture
"The Architecture persona created a full spec with OAuth 2.0, PKCE, and JWT — complete with a sequence diagram showing the entire login flow. Not prose. Real technical specifications."
A rendered Mermaid sequence diagram showing the OAuth login flow with token lifecycle, refresh rotation, and session management — created automatically as part of the architecture spec.
Code that follows the plan
"Implementation built the React components following the design spec — stat cards, booking table, availability chart. It wasn't just generating code; it was building what was specified, linked back to the requirements."
Code written in context — React components following the design spec, with the implementation task in the sidebar linking back to the specs and requirements it implements.
The running app
"It launched the dev server and there it was. A real dashboard — live clock, KPI cards, booking table with status badges, availability chart. Built from a conversation, with requirements, architecture, and tests behind it."
ReserveTable — 10 reservations, 38 guests, 5/20 tables available, $4,280 revenue. Status badges, availability chart, quick actions. Built with Crucible.
"I've tried building projects this complex with AI before. They always fall apart at scale — the AI forgets what it decided, contradicts itself, skips steps. Crucible didn't. The structure held. Every decision was traceable. And the system got better as I used it — the knowledge graph meant nothing was forgotten."
Want to try it?