← All posts

Under the hood ·

Your AI Agent Was Paying for 94,000 Tokens Before You Typed a Word. We Cut It to 7,000.

TRMNLY's agent used to send every tool's schema with every request. As of 0.1.103 it sends only the tools the conversation needs, so each turn is cheaper and leaves more room for you.

A tall stack of glass discs beside a much shorter stack glowing bright cyan, on a dark studio floor

Here's a cost most people never see. Every time an AI agent with tools answers you, the request includes a description of every tool it could call. In Anthropic's API, those definitions are part of the request, and they count toward your input tokens like any other text.1

TRMNLY has a lot of tools: terminals, panes, git, browsers, files, Docker, kanban boards and more. Until October 13, the agent sent the schema of every one of them with every request, which came to about 94,000 tokens of tool definitions per turn.2

As of TRMNLY 0.1.103, it sends the tools the conversation needs, and that brought tool definitions down to about 7,000 tokens.2

What that buys you

  • Cheaper turns. Fewer input tokens per request means a smaller bill, especially on long sessions where the same overhead repeats with every message.2
  • More room for the conversation. A model's context window is finite. Tool definitions you don't need were crowding out your code, your files and the chat history.2

The meter tells the truth now, too

The same release fixed the context meter. The percentage and the budget bar now read the same estimate, and both use the model's real context window instead of a 32k guess. Once the provider reports actual usage, the meter shows that. For Claude Code and Codex chats it shows what the CLI reports, plus their tool calls as cards and their reported cost. It turns orange past 75% and red past 90%.2

Since 0.1.108, long chats on Anthropic are also prompt-cached.3 Anthropic describes caching as a way to reuse a repeated prompt prefix, which lowers cost and latency.4 Sending fewer tools makes every request smaller, and caching makes the repeated parts of a long conversation cheaper.

Why it matters if you bring your own key

Since 0.1.98, a turn served by your own API key spends no TRMNLY credits. You're billed only by your provider.5 So every token we stop sending is money that stays in your own provider account.

Sources

  1. 1.Anthropic, "Tool use with Claude" (pricing and how tool definitions are sent). docs.anthropic.com
  2. 2.TRMNLY 0.1.103 release notes, "Agent requests shrink from ~94k to ~7k tokens of tools" and "An honest context meter" (October 13, 2026). trmnly.dev/updates/stable/appcast.xml
  3. 3.TRMNLY 0.1.108 release notes, "The agent sidebar, audited again: about 300 fixes" (October 17, 2026). trmnly.dev/updates/stable/appcast.xml
  4. 4.Anthropic, "Prompt caching." docs.anthropic.com
  5. 5.TRMNLY changelog, 0.1.98, "Your own API key no longer spends TRMNLY credits" (October 6, 2026).