open-design-mcp-agent-workflow
Build via Open Design MCP; drive agentic CLIs as sub-agents.
- License 许可
- MIT
- 最近更新
- 2026-10-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- open-design-mcp-agent-workflow
- tags
- [ "open-design", "mcp", "acp", "design-automation", "carousel", "pdf", "agent-adapter", "subagent" ]
- author
- Hermes Agent (captured from user workflow)
- license
- MIT
- version
- 1.0.0
- platforms
- [ "linux", "macos", "windows" ]
- description
- Build via Open Design MCP; drive agentic CLIs as sub-agents.
- related_skills
- [ "build-discipline", "writing-plans", "verification-loop" ]
02 资源文件 3
- 参考文档references/agent-cli-probe.md1.9 KB
- 参考文档references/pexels-fal-media-fallback.md1.8 KB
- 脚本scripts/render_carousel.py3.0 KB
03 SKILL.md 正文
Open Design MCP + Agent-CLI Workflow
How to use Open Design (github.com/nexu-io/open-design) as a toolbox that Hermes drives, and how to extend it with other agentic CLIs as sub-agents. Captured from a real end-to-end build: a travel carousel → 8 PNG cards + PDF, plus enumerating/harvesting a Telegram account's channels.
Mental model: two directions, only one we use
- MCP (Hermes → OD) — Hermes reaches INTO OD and calls its tools. OD is the toolbox; Hermes is the builder. This is the path we use. It works regardless of OD's run/execution state.
- Agent adapter / ACP (OD → Hermes) — OD dispatches tasks TO Hermes as an execution
engine. Useful only when OD's UI is driving. In practice the
start_runexecution path was broken in our setup (returns202thenstatus: failed), so we do NOT rely on it. We build by reading OD skill assets directly via MCP and rendering locally. Don't claim OD "executes" a skill for you — it didn't, and it's the reverse direction.
User's own framing (verbatim intent): "we can just call its mcp to build, we are the one building by using its mcps." — the builder is Hermes; OD supplies skills/templates.
Daemon (the MCP server host)
- OD runs a local privileged daemon. Keep it alive as a
systemd --userservice (open-design-daemon.service), pinned to the Node version its native addons need (Node 24.19.0 forbetter-sqlite3ABI 137). Auto-starts on boot. - Health:
GET http://127.0.0.1:7456/api/health→200. Also/api/skills(162 skills),/api/plugins(150),/api/projects. - The agent adapter is registered (
hermes.tsin the daemon's defs) and passeshermes acp --check, but that does NOT mean OD can execute runs — see caveat above.
MCP tool surface (the real interface)
- After a gateway restart that reloads MCP, the tools appear as
mcp_open_design_*. In a continuing session they may not be inline until a/reset— but you can always drive them directly via the daemon's stdio JSON-RPC (node cli.js mcp --daemon-url http://127.0.0.1:7456), which returns 22 tools including:list_skills,list_plugins,create_project,create_artifact,write_file,get_file,start_run,get_run,list_agents,collect_brief,confirm_brief. - Working build loop (proven):
create_project→write_file(base64) ×N → files land in OD's artifact/project store.start_runis the broken execution path; don't use it.
Building a deliverable (proven: travel carousel)
- Read the skill via MCP
list_skills; pick the template (e.g.card-xiaohongshufor a swipeable carousel,deck-swiss-internationalfor a deck,brandkitfor brand assets,canvas-designfor PNG/PDF). Skill assets live underskills/<id>/in the OD repo; copy the HTML/Tailwind scaffold. - Author the content (cards, copy, palette) as HTML/CSS. Keep the template's
structure (vertical
.deckof.carddivs, 1080×1440 / 3:4). - Media: OD's
fal-*image/video skills need a fal.ai key — which was LOCKED (HTTP 403, reason TOP_UP) in our run. Fallback that worked: Pexels via API key (api.pexels.com/v1/videos/search) for stock video/pics. Download real clips, wire<video>/<img>withposter=frames (videos don't autoplay in a static screenshot). - Render to upload-ready output (local, not OD): headless Chromium screenshots each
card → PNG; bundle PNGs into a PDF with PIL. See
scripts/render_carousel.py. - Write deliverables into OD via MCP
write_fileso they show in the OD UI.
Rendering cards → PNG + PDF (the local step)
OD skills produce HTML/CSS, not image files. To get postable assets:
- Per-card HTML: extract each
.carddiv into its own minimal HTML file (the?card=Nquery-param isolation trick FAILS under headless Chrome — it ignores the param; you get 8 identical screenshots). Building one file per card is the reliable way. - Screenshot:
chromium --headless --no-sandbox --window-size=1080,1440 --screenshot=out.png file:///path/card-N.html. Crop to 1080×1440 if the viewport differs. - PDF:
PIL.Image.save(pdf, save_all=True, append_images=[...])→ one page per card. - Video cards: put a
poster=still (extract withffmpeg -ss 2 -i clip.mp4 -frames:v 1) so the static card shows footage; the live<video>stays for the HTML deck. - Gotcha: a "full page"
--screenshotonly captures the first viewport, not the whole scrollable deck — screenshot per card, never the whole page.
Driving other agentic CLIs as sub-agents
Verified on the host (find via command -v): pi, kilo, kiro, antigravity,
cursor, gemini, codex, claude, qwen are present; opencode/devin/grok/
deepseek/vela are NOT. To test one as a sub-agent:
- Run it non-interactively with a
-p/execprompt; check auth. Onlypireturned a clean authenticated answer (pi -p "..."→ works, has read/bash/edit/write tools).claude/geminisaid "not logged in";codexerrored on skill YAML;qwenhung;antigravitylaunched a browser and hung. Sopiis the one proven sub-agent; the rest need auth or are broken in this env. Don't claim the others work. pican read a repo, analyze code, and propose commands — it successfully diagnosed tgforwarder's forward logic and produced the correct (after flag-fix) commands.
Pitfalls (this user's triggers — avoid)
- Claiming OD "ran the skill" — it didn't via start_run; we built from its assets.
- Per-card screenshot via query param — produces identical images; build one file per card.
- fal.ai assumed working — it was LOCKED (TOP_UP). Use Pexels stock as the fallback.
- Claiming claude/codex/gemini are usable sub-agents — only
piwas verified. - "Excuses / it can't be done" — user mandate is find a way. Diagnose, then deliver.
Verification
- Daemon
HTTP 200on/api/health. - MCP
tools/listreturns 22 tools. - Deliverable: open the HTML deck in a browser; confirm 8 distinct PNGs (different file sizes = different content) + a multi-page PDF; visually check 1-2 cards.
Reference files (session-specific detail)
references/pexels-fal-media-fallback.md— exact Pexels search/download calls and the fal.ai 403 TOP_UP finding, so a future session doesn't re-discover the lock.references/agent-cli-probe.md— the non-interactive probe results for each CLI (pi works; others need auth/broken).scripts/render_carousel.py— the Chromium+PIL render script (per-card HTML → PNG → PDF).
04 同仓库更多 Skills
blocked-page-recovery
Use when a fetch fails: 403/429, paywall, WAF, bot wall.
gemini-vision-fallback
Vision fallback: Gemini REST API if vision_analyze blocked.
twitter-cli
Use twitter-cli for ALL Twitter/X operations — reading tweets, posting, replying, quoting, liking, retweeting, following, searching, user lookups. Invoke whenever user requests any Twitter interaction.
prism-scan
Structural analysis through dynamically generated cognitive lenses. Generates the optimal analytical lens for the specific code/artifact, then executes it. Finds conservation laws, structural invariants, and concrete bugs that vanilla analysis misses. Use on any code file, system design, or text artifact.
prism-reflect
Constraint transparency: analyzes an artifact structurally, then analyzes what its own analysis concealed. Produces a conservation law AND a constraint report showing what was maximized, what was sacrificed, and what to investigate next. The only AI skill that knows what it can't see.
prism-full
Full Prism: multi-pass structural analysis with mandatory adversarial self-correction. Designs custom analytical passes, executes them with chaining, then attacks its own findings before synthesizing. Use for maximum depth on important code or artifacts.