Agent and chatbot token bloat

A chatbot with a short memory is not the same bill as a coding agent that pastes every command. Tokand is built for the second case: tool traces that get resent on every turn.

What gets billed twice

Agents call tools, then send the tool output back to the model so it can continue. The next turn often includes that output again, plus the one before it. The useful fact is a few lines. The rest is logs, diffs, and status text. That is the junk Tokand is meant to shrink. Facts stay. Stale tool turns can be dropped. If the tool result already answers the question, the model call can be skipped.

Stacks people ask about

Cursor — agent runs that reread the repo. Cursor guide.
Codex and GPT agents — command output piled into the next prompt. Codex guide.
OpenClaw and Hermes — harnesses that already speak a base URL. OpenClaw, Hermes.
Chatbots — a one-paragraph prompt with no tool output should bypass the filter and record zero saved.

Same answers, less leftover text

Tokand is not RTK and not Caveman. An RTK-style filter shortens command and tool output and keeps the facts. A prose rewrite can change answers, so that path is off by default and, on Aggressive+, only touches older user and assistant turns. The newest message, the system prompt, and tool results stay as you sent them.

If the time spent filtering would cost more than the prefill it removes, the call goes around the filter.