Introducing Agent Usage Credits
Agent no longer requires an API key. starting today, any account with developer edition can build plugins for free with a model we're hosting on your behalf: DeepSeek V4.1 Flash via OpenRouter.
new accounts start with enough credits to build ~5 plugins, and TRMNL+ members get free credits every month, forever. you can also top up an account with $5, $10 or $20 one-time, and connecting your own model is still supported.
How usage credits work
see your live balance at the top of Account › AI Agent. this meter is also visible in the bottom of the Agent panel while you build.
usage credits are spent before tapping into your own model or topped up cash. we add 10% and 30 cents to maintain this service, and this fee is transparent at checkout.
You add | You pay |
|---|---|
$5 | $5.80 |
$10 | $11.30 |
$20 | $22.30 |
but again - if you already have an OpenRouter or Anthropic key, nothing changes. add it below Bring your own key and the Agent uses your stack instead.
What a plugin actually costs
we ran several tests with DeepSeek v4. here are results from a one-liner i gave it yesterday (typo included): "make an AirGraident plugin, here's their API: <link>"
the Agent read the docs and got live data flowing from a public weather station. it asked if i wanted to make it flexible for personal and public monitors, to which i said yes, and boom: custom form fields appeared.
this build took 70 API calls, totaling 2.6 million tokens in and 43,000 out. the token spend was $0.0393.
but AirGradient is just 1 example. this week, people have built plugins for anywhere from 0.6 cents to 10.6 cents, with a median cost of 4.4 cents. with these benchmarks, even a $5 top-up affords ~100 custom dashboards without writing a single line of code.
meanwhile, here's what a typical Agent build costs using other supported models:
we'll continue to support leading models, but our Framework design system and Agent Skills make DeepSeek an obvious choice for most people and most custom plugins.
Why it's so cheap
we've built or helped build thousands of ePaper dashboards. nearly everything we know about the subject is both open source and well documented. more specifically...
cheap tokens. with OpenRouter, DeepSeek V4.1 Flash costs ~$0.14 per million input tokens and $0.42 per million output tokens. tokens it already saw in the prompt cost $0.004 per million. since DeepSeek reads images too, which our Agent uses to generate and compare screenshots against Framework CSS expectations.
lean prompt. invoking the Agent includes a set of battle-tested instructions, totaling 168 KB of text, roughly 40,000 tokens, which the provider caches. from my AirGradient test build, 97% of the input tokens were cache reads. across all TRMNL DeepSeek V4.1 Flash chat so far, cache reads account for 87% of input tokens. without this in place, the same plugin would have cost 10x or 39 cents (vs 3.9 cents).
knows how to meet ladies (HTML). with models now being used to cure cancer, they're more than capable of writing HTML and JavaScript. the prompts also make it aware of our {{ liquid }} templating engine and other core features.
loops. as Agent builds plugins, it pulls reference examples from the public marketplace. it also screenshots layouts and fixes them automatically.
Bring the same skills to your own agent
the instructions and design system our Agent runs on are open source:
usetrmnl/trmnl-agent-skills.
- Claude Code:
/plugin marketplace add usetrmnl/trmnl-agent-skills - Cursor, OpenCode, Codex, Gemini CLI, GitHub Copilot: drop-in files, see the repo's install guide
pair our Agent Skills with TRMNL's new MCP server and your agent can build plugins, update settings, and take screenshots just like ours. as of this writing, TRMNL's MCP server has answered 174,532 tool calls from 749 people across 1,175 plugins.
Get started
open any private plugin's markup editor, click Agent and describe what you want. if you build something that could help others, consider publishing it and getting paid.
