Where the tokens go
An AI coding agent burns tokens two different ways, and they have almost nothing to do with each other.
The first is the conversation. Every turn re-sends a growing pile of context — files it already read, the system prompt, earlier turns, the plan. The model bills you for all of it, again, on every request.
The second is command output. The moment an agent runs
git status, npm test, grep or
kubectl get, the full, verbose result gets pasted straight back into
the context — borders, timestamps, stack traces, hundreds of lines of log. That
output wasn't in the conversation a second ago; the agent's own tools just put it
there.
headroom was built for the first problem. RTK is built for the second. Slipstream now runs both.
headroom — compressing the conversation
headroom is the engine Slipstream launched with, and it stays the core. It runs as a local proxy: your agent points its model endpoint at Slipstream instead of straight at the provider, and every request passes through headroom first. It rewrites the context — dropping redundancy, trimming what the model has already seen — before forwarding a much smaller request upstream. Same answer, 60–95% fewer tokens.
It works at the API layer: it sees the whole conversation on its way to the model and compresses it as one. That's the right place to catch re-sent context, but by the time a request reaches headroom, a noisy command result is already part of it.
RTK — compressing the commands
RTK works
one step earlier, at the shell layer. When your agent is about to
run a Bash command, RTK rewrites it — git status becomes
rtk git status — so the command runs through RTK. RTK knows
the shape of 100+ common dev commands and returns a compact, structured version of
their output: the same information, with the borders, repetition and dead weight
stripped. Often 60–90% smaller, in under 10 ms.
The key point: RTK trims that output before it ever enters the context. It's not compressing the conversation — it's stopping noise from becoming part of the conversation in the first place.
Why run both
Because they're not the same tool pointed at the same problem. They sit at different layers and save tokens from different sources, so their savings add up rather than overlap.
| headroom | RTK | |
|---|---|---|
| Layer | API · the conversation | Shell · command output |
| Compresses | context sent to the model | output of commands the agent runs |
| Mechanism | local proxy rewrites each request | hook rewrites Bash commands |
| Best at | redundant, re-sent context | noisy dev output (git, tests, logs) |
| When it acts | on the way to the model | before output enters context |
With both on, RTK trims command noise at the source and headroom compresses whatever conversation remains. One reduces what gets created; the other reduces what gets re-sent. The dashboard now splits your savings by source — Compression, Prefix cache, and Commands — so you can see each layer's contribution.
A first implementation
We want to be straight about what this is: a first pass. RTK support landed in v0.1.5, it's off by default, and you turn it on per-machine.
RTK reports its savings in tokens, not dollars, so the figure you see on the dashboard under "Commands" is an estimate — we convert those tokens to a dollar value using a representative input-token price. The real combined effect of running both engines depends heavily on how command-heavy your agents are, which models you use, and how your provider prices cached input.
We don't yet have enough real-world data to put a confident number on total combined savings. As more people run both engines, we'll sharpen the estimate and share what we find here. Treat the Commands figure as directional for now, not exact.
Turning it on
RTK is bundled with Slipstream — there's nothing extra to install. In
Settings → Command compression, flip
Trim command output (RTK). Slipstream wires it into Claude Code and
makes the bundled rtk available on your path.
A few things worth knowing before you do:
- Restart your agent. The hook loads when a Claude Code session starts, so restart any running sessions after toggling.
- It rewrites Bash commands only — Claude Code's built-in Read/Grep/Glob tools bypass it.
- Claude Code first. This release wires RTK for Claude Code; other agents follow.
- macOS & Linux. On native Windows RTK falls back to a CLAUDE.md-based mode without the automatic rewrite (WSL recommended).
The full walkthrough — with every caveat — is in the manual. New to Slipstream? Download it and the conversation layer (headroom) is on out of the box; add the command layer whenever you like.