HN 표시: Scryer – AI 에이전트를 위한 시각적 아키텍처 모델링

MDD for AI agents. Scryer is a visual C4 modeling tool where you and your AI agent work from the same architecture model. Features • Getting started • MCP server • Building from source • Docs demo.mp4 AI agents can write code, but what you describe and what gets implemented tends to drift. Scryer gives you a shared visual model: you edit it in a drag-and-drop editor, the agent reads and modifies it through MCP. Once the model looks right, the agent generates code from it — get_task feeds work one unit at a time with dependency ordering, inherited contracts, and progress tracking. Works with Claude Code and Codex out of the box. Any agent that supports MCP can read and write models. Agents that support ACP can also be spawned by Scryer for automated sync. Opinionated C4 hierarchy (system, container, component, operation/process/model), typed relationships, behavioral flows, contracts. - C4 Architecture Diagrams - Drag-and-drop editor for systems, containers, components, and operations. Drill down through levels. - Code-level nodes (operations, processes, models) show in a compact list view. - Behavioral Flows - Model user journeys, data pipelines, deploy sequences. Supports branching and decision points. - Flows serve as integration test specs — link them to test files via source mapping. - Contracts - Expect/ask/never rules that tell AI agents how to implement your code. Inherited down the hierarchy. - Expect items have pass/fail flags that control when a node can be marked "verified". - Status Tracking - Four statuses: proposed (planned), implemented (code exists), verified (contracts satisfied), vagrant (discovered during sync). - During implementation, agents mark nodes as implemented. "Verified" is a separate step — the implementation must be complete (no stubs or TODOs), existing tests must pass, and all expect contract items must be satisfied. - Source Mapping - Link architecture nodes to files in your codebase with file patterns and line ranges. - Click to open in your editor. - MCP Server - AI agents connect to read, modify, and build from your architecture model in real-time. - Drift Detection & Sync - Scryer tracks when source files change relative to the model. When drift is detected, click sync to have Scryer spawn the connected agent to update the model. - AI Advisor - Optional LLM-powered review that flags structural issues in your diagrams. Supports OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, and Ollama. - Implementation Workflow get_task gives AI agents one piece of work at a time, ordered by dependencies, with contracts inherited from parent nodes. Build, mark implemented, repeat. - AI Tool Setup - Detects Claude Code and Codex, writes MCP config and auto-approve permissions for your project. Download the latest release for your platform from the releases page. - Link your project directory in the app and enable AI tool integration when prompted (or run scryer-mcp init ) - Tell your AI agent: "Use scryer to model this project's architecture" - The AI calls MCP tools — nodes appear in the visual editor in real-time - Review, drag things around, rename, remove, restructure - Tell the AI: "Implement this model" - The AI builds each piece one at a time, marking nodes as implemented as it goes - When you're satisfied, ask the AI to verify: check for stubs, run any existing tests, and confirm contract items pass before marking nodes as verified As you work on code, Scryer detects when source files drift from the model. Click the sync button to have Scryer spawn your agent to update the model. Scryer is built to work with Claude Code and Codex first. - MCP (Model Context Protocol) — how agents read and write architecture models. Required for any agent integration. - CLI spawning — how Scryer launches agents for automated sync. Claude Code is spawned via claude -p (uses your subscription), Codex viacodex exec (uses your API key). Both get the Scryer MCP server attached automatically. - ACP (Agent Client Protocol) — for agents that implement the full ACP handshake (e.g. via claude-agent-acp). Scryer falls back to ACP if a {name}-acp binary is found on PATH. When an agent connects via MCP, Scryer captures its identity from the protocol handshake. When sync is triggered, Scryer resolves that identity to a binary and launches it with the right flags. Claude Code and Codex are mapped automatically. For other agents, Scryer tries ACP conventions. The MCP server lets AI agents read and modify your architecture models. It ships bundled with the desktop app. Link a project directory in the app and click "Enable" on the prompt, or run scryer-mcp init from the command line. Both detect installed AI tools and write config: - Claude Code — .mcp.json + read tool auto-approve in.claude/settings.local.json - Codex — .codex/config.toml Existing config files are preserved — only the scryer entry is added or updated. If you prefer to configure MCP manually, add Scryer to your project config: Claude Code (.mcp