Tokand, in public

Tokand (TOK&) sits in front of LLM and agent API calls and cuts billable tokens. Same models. Same answers. If Tokand is down, the call goes around us.

What it does

Filters tool and command output so facts stay and leftover logs do not.
Drops stale tool turns that the next call would resend.
Can skip the model when the tool result already answers the question.
Leaves a short prompt with nothing to cut unchanged, and records zero saved.
Skips the filter when the added wait would cost more than the prefill it removes.

How you install it

Base URL. https://api.tokand.com/v1 on any OpenAI-compatible client.
One call. tokand(request, sender). Tokand shrinks the request. Your code still sends it.

Tokand does not wrap the model SDK. Steps: quickstart.

Tiers

Safe — keep the answer path strict.
Measured — more of the junk goes, still aimed at the same answer.
Aggressive — stronger cut of tool leftovers.
Aggressive+ — everything aggressive does, plus a filler-phrase cut on older user and assistant turns. The latest user message, the system prompt, and tool results stay verbatim. Off unless you select this tier.

Agents

Claude, ChatGPT, Claude Code, Codex, Cursor, OpenClaw, Hermes, and anything that accepts a base URL. Hub: all agents.

What these pages never include

Customer prompts, API keys, organization names, invoices, usage from a real account, or dashboard data. Contact is hello@tokand.com.

Writing for search and agents: blog. Machine summary: llms.txt.