🪴 Brain

Memegen

4 min read

Memegen

🗒️ jacebrowning/memegen is the open-source service behind memegen.link: a URL-based meme API. You pick a template id, put the lines in the path and get an image back, no account and no key. MIT, about 1.75k stars, last push 2026-10-15 (mostly dependency bumps), with an agent-oriented URL guide added in May 2026. I use it as the hosted renderer in the Meme Generation for Coding Agents design, because it is the only hosted option that behaved in every test.

Description

  • Two ways to call it. A GET URL built by the agent (/images/<template>/<line1>/<line2>.png, with _ for a space, ~q for ?, ~s for /, ~n for a newline) or POST /images with JSON {template_id, text: [...], extension: "png"}, which returns a url. A POST took about 0.2 s and the cold render about 1.3-1.5 s; GIF and WebP take 6-10 s. Output formats are png, jpg, gif and webp; some templates animate.
  • Templates. 209 of them, listed at GET /templates/. The metadata is thin: keywords are empty on 158 of 209 and mostly franchise names on the rest (“The Office”, “Futurama”), not meanings. POST /images/automatic matched an English prompt well but returned an unrelated image for Polish text and for a “you’re absolutely right” prompt.
  • Polish works. The default font is Titillium Web Black (not Noto Sans Bold, as I first assumed); it, Impact and Noto Sans all rendered “Zażółć gęślą jaźń” correctly, and the forced uppercase is correct for Polish. Templates force uppercase, with no per-request case control.
  • Terms to know. Requests without a key are watermarked, may be rate-limited and may be used as training data; the hosted service also forwards caption text and referer to the maintainer’s tracking service (from the source). An API key comes with a 10 USD per month GitHub sponsorship (X-API-KEY header, not the URL) and lets you set your own watermark.
  • Self-hosting is documented (make run, a Heroku deploy and a Containerfile), though there is no official Docker image (#726) and the instructions lean to Heroku, macOS and Linux.

Download or use

Nothing to install for the hosted API. Minimal Claude Code route: a skill that fetches https://api.memegen.link/templates/ once (or ships a curated local catalog), has Claude pick the template and write the lines, builds the URL, downloads the PNG to a local file and opens it with start file.png. Use Node fetch with UTF-8 bodies: a Polish JSON body pasted inline through Git Bash curl arrived garbled. %, #, /, ? and _ in captions need the escape rules, and % and # only worked after memegen’s 301 redirect.

Reasoning for

I want one renderer that needs no native binary and no account, for non-sensitive captions. Memegen is that, with three caveats: the watermark, the training-data clause, and the fact that every hosted caption can end up cached for weeks. Hence the rule in my design: use it only as a fallback behind a local renderer, and never for captions that name a client or a project. The thin template metadata is the reason the design keeps its own catalog with a “when to use” line per template.

⚠️ Known issues from the research: custom backgrounds need a public URL, possible Pillow memory growth on small dynos (#860), and no case control. Several skills wrap memegen and inherit all of this: Meme Factory, makememe (which has no API-key support, so even sponsoring would not remove its watermark) and memelink-cli.

Alternatives considered

  • Imgflip API: free get_memes (about 100 templates) and caption_image, watermark on the free tier; auth moved to Bearer keys around 2026-09-30 while wrappers still use passwords.
  • Local rendering with Satori or a Pillow script: no watermark and no data leaves the machine, at the cost of owning a template catalog. This is my primary route.
  • OpenClaw Meme Maker: a curated registry plus local rendering, Imgflip as an option.
  • Memesio and other hosted MCPs: forced watermarks, anonymous operators, brittle.

Resources

Template: tool

Graph View

Backlinks