What is fireworks-tech-graph?
fireworks-tech-graph is an Agent Skill, a SKILL.md plus Python scripts, that lets Claude Code or Codex draw technical diagrams from a plain-English or Chinese description. The agent classifies the request, writes a versioned JSON description of the diagram, and the bundled generator routes the arrows and renders SVG, a 1,920-pixel PNG, an offline interactive HTML page or an animated GIF.
It covers 14 diagram types, including the UML set (class, sequence, state machine, deployment and others) and ER diagrams, plus ready patterns for AI systems such as RAG, multi-agent setups, tool-call flows and Mem0-style memory. There are 12 visual styles, 11 driven by the generator and one written freely by the model, and icons for 40+ products.
It is MIT-licensed, written mainly in Python, by Brad Zhang (yizhiyanhua-ai). Version 1.2.0 shipped on 17 July 2026; the last code change landed on 5 September and the repo had about 11,800 stars at last check.
Is fireworks-tech-graph worth it?
fireworks-tech-graph is worth installing for one reason: it does not trust the model's first drawing. The skill has the agent write a schema-validated intermediate description, a deterministic planner routes the edges, and the result has to pass structural checks (XML, markers, overlaps, crossings) before the agent reads back the PNG and makes at most two targeted fixes. The showcase contract asks for zero crossings, no more than two bends per edge and at least 40 pixels between nodes. When the agent cannot read images, it reports `visual_review: skipped` instead of claiming a pass.
The output is aimed at slides, blog posts and design reviews, not living documentation. Diagrams come out as SVG and PNG, which look far better than Mermaid but cannot be diffed or edited as text in a pull request. Style 9 produces C4 container reviews and styles 10 to 12 cover multi-region deployments, event topologies and golden-signal dashboards, which are closer to what platform teams actually draw. Two of the styles imitate Claude's and OpenAI's house look, which is fine for internal decks and questionable on a public page.
Installation has rough edges. The README warns that a bare `npx skills add` of the repo can pick up the wrong SKILL.md, so you have to give the nested path, and the npm package (1.0.4) lags the GitHub release. PNG export needs cairosvg or rsvg-convert, and GIFs need FFmpeg, Chrome and Puppeteer installed beside every copy of the skill.
Maintenance is one person's. There are 12 contributors, one open issue, and the only commit since 5 September is a README sponsor edit; the README also advertises the author's paid consulting. Compared with Anatomy, which builds one interactive explanatory figure over a long session, this skill is quicker and more conventional: boxes, arrows and swim lanes, done properly.
Where it earns its place
A staff engineer preparing a design review
Draw the proposed checkout architecture without an afternoon in draw.io
Ask Claude Code for a C4 container diagram in style 9 with named services and protocols; the skill renders SVG and PNG and reports validation results. Cost is one agent session; edits mean re-prompting rather than dragging boxes.
A developer advocate at an AI infrastructure company
Illustrate a RAG or multi-agent pipeline for a launch blog post
Use the built-in RAG or multi-agent pattern with product icons and export a 1,920-pixel PNG plus an animated GIF for social posts. The GIF path needs FFmpeg and Puppeteer installed first.
An SRE lead documenting an event-driven system
Show topics, consumers and dead-letter queues in one picture
Style 11 draws topics as rails and processors as stations, with a declared DLQ; the offline HTML export gives readers pan and zoom without hosting anything.
Before you commit
- Setup
- `npx -y skills@1.5.17 add yizhiyanhua-ai/fireworks-tech-graph/skills/fireworks-tech-graph --agent codex claude-code -g -y --copy`, or a git clone into ~/.claude/skills or ~/.agents/skills. Needs Python 3.9+ and cairosvg; `fireworks.py doctor` checks the optional GIF dependencies.
- Integrations
- Claude Code and Codex through the Agent Skills format; a unified CLI (render, validate, check, export-html, animate) that also works without an agent; renders with cairosvg, rsvg-convert or headless Chrome.
- Data and privacy
- The README describes no hosted service: rendering and checks run locally, and the description you type goes only to the model your agent uses. MIT licence, so commercial use is unrestricted.
- Cost at scale
- Free. The cost is agent tokens, including up to two correction passes and an image read-back per diagram; no per-diagram token figures are published.
- Limits
- No text-based diagram source for docs-as-code workflows; style 8 is written freely by the model and skips the generator; rsvg-convert may drop CSS, so PNGs look best with cairosvg.
What's good
- Validator and PNG read-back before a diagram counts as done
- 14 diagram types including full UML and ER
- AI-system patterns (RAG, multi-agent, tool calls) built in
- SVG, PNG, offline HTML and GIF outputs
What's not
- Image output, not text you can diff in a pull request
- Nested install path required; npm package lags
- GIF export needs FFmpeg, Chrome and Puppeteer
- Single maintainer; little code change since September
Best for
Engineers and technical writers who need polished architecture diagrams for slides, posts and design reviews from inside Claude Code or Codex.
fireworks-tech-graph alternatives
fireworks-tech-graph: common questions
- Is fireworks-tech-graph worth it?
- A diagram skill that checks its own geometry before handing over, ideal for slides and posts, though Mermaid still wins for diagrams that live in the docs. We rate it 3.9 out of 5.
- How much does fireworks-tech-graph cost?
- Free and open source (MIT). Runs on the agent subscription or API you already use.
- What are the downsides of fireworks-tech-graph?
- Image output, not text you can diff in a pull request; nested install path required; npm package lags; gIF export needs FFmpeg, Chrome and Puppeteer; single maintainer; little code change since September.
- Who is fireworks-tech-graph best for?
- Engineers and technical writers who need polished architecture diagrams for slides, posts and design reviews from inside Claude Code or Codex.
- How much does fireworks-tech-graph cost for a team or in production?
- Free. The cost is agent tokens, including up to two correction passes and an image read-back per diagram; no per-diagram token figures are published.
- Is fireworks-tech-graph safe for company data?
- The README describes no hosted service: rendering and checks run locally, and the description you type goes only to the model your agent uses. MIT licence, so commercial use is unrestricted.
- What are the best fireworks-tech-graph alternatives?
- The closest alternatives we have reviewed are Anatomy, Anthropic Skills and Impeccable.
Answers checked Oct 26, 2026.