Wrapping a single run

wrap is the most reliable way in, and the one to start with.

tokunseba wrap claude

It launches the tool with the proxy's base URLs set on that process, for that run only. Nothing is written to a settings file, nothing is added to your shell profile, and when the tool exits nothing is left behind.

Why it is more reliable than initCopy Anchor Link

Writing a tool's settings file can be overridden by an exported environment variable, and several launchers export one. An exported ANTHROPIC_BASE_URL pointing at the provider beats anything init writes, silently.

Wrapping sets the variables on the tool's own process, so nothing inherited can win.

Passing the tool its own argumentsCopy Anchor Link

Everything after -- goes to the tool:

tokunseba wrap codex -- --model gpt-5
tokunseba wrap aider -- --no-auto-commits src/

Anything not in the listCopy Anchor Link

tokunseba wrap-any -- my-agent --flag

exports every tokunseba base URL and runs the command. Use it for a tool that is not in the known list, or for a whole shell:

tokunseba wrap-any -- zsh

Everything started from that shell is routed, and nothing outside it is.

Compressing one command's outputCopy Anchor Link

A different thing with a similar shape, and often the highest-value change you can make:

tokunseba run -- pytest -q
tokunseba run -- npm install
tokunseba run -- cargo build 2>&1

Runs the command, prints its output already compressed, and exits with the command's own exit code. No proxy is involved at all.

If your agent can run shell commands, telling it to prefix long-running commands with tokunseba run -- stops the output being large in the first place, rather than shrinking it after it is already in the transcript.

Anything left out is behind a handle, exactly as it would be through the proxy:

tokunseba expand h_db4d5709