Reladraw is a TypeScript tool that takes a plain-text file of relative placement statements — "node store right of app level with app" — and renders an SVG. No coordinates anywhere in the source. No auto-layout algorithm choosing arrangements for you. You state relationships, and a constraint solver finds the tightest layout that satisfies all of them. The result is deterministic: same file, same picture, every time. The pitch is sharp and honest. Mermaid and Graphviz are superb when you want a machine to arrange things. Draw.io and Excalidraw are superb when you want to place every pixel by hand. Neither works well when you have a specific picture in your head but don't want to manage coordinates. Reladraw fills that gap with a grammar of relative positions — above, below, right-of, between — resolved by longest-path computation on each axis. One answer, no search, no heuristics. The agent story is where this gets interesting. An LLM emitting Mermaid is guessing at a layout that an algorithm will settle later, with no way to verify the result without rendering and looking. An LLM editing draw.io XML is doing pixel arithmetic blind. Reladraw's stated-placement model means the intent is readable in the source file: the agent can re-read its own output and confirm that the database is below the API and the four machines hang off the hub. That's a genuine structural advantage for agentic workflows. The implementation is clean: zero runtime dependencies, a parser, a deterministic resolver, and an SVG renderer. The 44-statement example file produces a network architecture diagram that matches a hand-drawn draw.io equivalent. The resolver works by treating each axis as a set of minimum-distance constraints and finding the tightest satisfying arrangement via longest paths — elegant, and it means non-overlap comes free without anyone declaring it. What's not built yet matters. Edge routing around nodes doesn't exist — edges are straight lines that cut through whatever's in the way. The diagnostics system that would tell you "hub overlaps laptop1" isn't implemented. The icon and shape library is a closed set. Typography is visibly rough. These aren't nitpicks; they're the difference between a proof of concept and a tool you'd use on a real architecture diagram tomorrow. The design philosophy is rigorous to the point of being a thesis statement. The engine is forbidden from choosing arrangements — it may affect distances but never relationships. Auto-layout is refused on principle: a picture chosen by an algorithm is not predictable from its source, and predictability is the entire point. That's an opinionated stance that will either attract a devoted user base or limit adoption, depending on how much of diagramming turns out to be "I know where things go" versus "figure it out for me." At v0.5.0, this is early-stage open-source work with a clear vision, a working pipeline, and significant gaps. The language isn't stable. The skill-installation system for AI agents (via npx skills add) is a smart distribution move. Whether reladraw earns its middle ground depends on whether the grammar can express enough real diagrams — the authors acknowledge their benchmark contains arrangements the grammar can't yet handle. That's the honest test ahead.
