What Slipstream does
An AI coding agent re-sends a lot of the same context on every turn: files it already read, stack traces it already saw, logs, lock files, history. The model bills you for all of it, every request.
Slipstream sits in the middle as a local proxy. When your agent sends a request, it passes through Slipstream first, which removes the redundant parts before forwarding it to the model provider. The agent gets the same answer; the request is 60–95% smaller. The compression engine is headroom, an open-source (Apache 2.0) project; Slipstream is the desktop app that runs it with zero setup and shows you, in dollars, exactly what it saved.
- Runs entirely on your machine — no account, no cloud, no servers.
- Your agent keeps its own API keys and its own connection to the provider.
- Works with Claude Code, Codex, Cursor, Aider and GitHub Copilot, plus any other OpenAI-compatible tool.
- Free for 30 days, then a few dollars a month on the honor system.
1 · Install
Slipstream is a regular desktop app — download, open, done. Pick the build for your system. macOS 13+ and most current Linux distributions are supported. A signed Windows build is on the way.
On the download page you'll find macOS (Apple silicon) and
macOS (Intel) .dmg files and a Linux
.AppImage. Not sure which Mac you have? Apple menu → About This Mac:
"Apple M1/M2/M3…" means Apple silicon, "Intel" means the Intel build.
macOS
-
Open the .dmg and drag Slipstream to Applications
The standard install — drop the app icon onto the Applications folder shortcut.
-
Launch it from Applications
Use Launchpad or Spotlight. The app lives in your menu bar, not the Dock.
If you see "Slipstream cannot be opened because the developer cannot be verified," right-click (or Control-click) the app in Applications and choose Open, then confirm once. Alternatively: System Settings → Privacy & Security → scroll down → Open Anyway. You only do this the first time.
Linux
-
Make the AppImage executable
Either right-click → Properties → "Allow executing file as program," or in a terminal:
chmod +x Slipstream_*.AppImage -
Run it
Double-click the AppImage, or run it from the terminal. It appears in your system tray.
2 · First launch
The very first time you open Slipstream, it sets up a private runtime for the compression engine. You'll see four short steps:
-
Downloading Python
A self-contained Python — not your system one, and it never touches an existing install.
-
Creating a private environment
An isolated folder that belongs only to Slipstream.
-
Installing headroom
A pinned, known-good version of the compression engine.
-
Verifying everything works
A quick self-check, then the app opens.
This takes about a minute on a cold machine and only happens once. It needs a network connection that first time, to fetch the runtime. After that, Slipstream runs offline — the only network traffic is your agent's own calls to its model provider.
"No Python required" means you don't have to install or manage Python yourself. Slipstream provisions its own and keeps it in its own directory, so nothing on your system changes.
4 · Starting the proxy
The proxy is the part that does the work. Nothing is compressed until it's running. Turn it on from the menu-bar switch, or from Start proxy on the dashboard. By default it listens at:
http://localhost:8787/v1
That's the address your tools point at. Most agents are wired up for you (next
section); this URL is for anything you connect by hand. You can change the port in
Settings → Proxy → Port if 8787 is taken.
To have it ready whenever you are, turn on Start proxy when Slipstream opens and Launch Slipstream at login in Settings. Then it's always there, quietly, and savings begin the moment you start coding.
5 · Connecting your agents
Open the full window and go to Agents. Slipstream detects the supported coding agents you have installed and lists them. Connecting one is called wrapping: Slipstream edits that agent's own config so its traffic flows through the proxy automatically.
- Agents with a switch — flip it on. That's the whole step.
- Agents with a "Snippet" button — these wrap per session. Click Snippet, copy the line, and use it where that tool reads its config or when you launch it.
-
Anything else — any OpenAI-compatible tool works by pointing its
base URL at the proxy (the
http://localhost:8787/v1address above).
An agent only reads its config when it starts. So a session that's already running keeps its old connection until you restart it — flip the toggle, then restart the agent (or open a new session) to route it through Slipstream.
Wrapping is fully reversible. Turn the switch off and Slipstream restores whatever routing was there before.
Claude Code
Flip the toggle, then restart Claude Code. Wrapping Claude Code also installs a small startup hook that wakes the proxy when a session begins — so a freshly started session never points at a dead port, even if Slipstream wasn't open yet.
Codex, Cursor, Aider, GitHub Copilot
Same idea: flip the toggle in the Agents list (or copy the Snippet if that agent shows one instead), then restart the tool. Each row shows its current state — "Routing through Slipstream" once it's wrapped, "Not wrapped" when it isn't.
Any other OpenAI-compatible tool
Set the tool's API base URL (sometimes called "OpenAI base URL," "endpoint," or
"OPENAI_BASE_URL") to the proxy address. Keep using your own provider
API key as normal — Slipstream passes it straight through.
OPENAI_BASE_URL=http://localhost:8787/v1
6 · Reading the dashboard
The Overview is the home of the full window. Once your wrapped agents start making requests, the numbers fill in. Use the range switcher in the top-right — Today, 7d, 30d, All — to change the window everything is measured over.
The four headline numbers
| Card | What it means |
|---|---|
| Tokens processed | Everything your agents sent through the proxy in this period — saved plus what still went to the model. |
| Tokens saved | The tokens Slipstream removed before they reached the model. |
| Cost saved (est.) | Those savings in dollars — compression plus any prefix-cache discount. Estimated from the model prices in Settings. |
| Token reduction | Saved as a share of processed — the headline percentage. |
Each card carries a small sparkline and a delta versus the previous period (for example, "↑ 18% vs previous"), so you can see whether savings are trending up.
The panels below
- Token savings over time — a bar per day, split into "saved" and "spent," so the shape of your usage is visible at a glance.
- Breakdown — where the dollars came from, split by source (rewriting context vs. provider prefix-cache) and by model.
- Recent activity — the latest agent sessions: which agent, which model, when, and how much each saved. It refreshes on its own.
When it looks empty
If the proxy is running but you haven't sent a request yet, the dashboard says it's listening and waiting — point a wrapped agent at it and the first numbers land here. If the proxy is stopped, you'll see a Start proxy button instead. Cost figures stay at zero until at least one request has gone through.
7 · Compression & settings
Everything is in Settings, and changes apply immediately — the proxy restarts itself when it needs to. No file editing, no JSON.
Compression mode
Two ways to balance raw savings against your provider's caching:
- Max savings — rewrites context aggressively for the biggest token reduction. The default, and usually the right choice.
- Cache friendly — freezes earlier turns so your provider's prefix cache keeps hitting. Pick this if your model bills cached input much cheaper and you want those cache hits preserved.
Command compression (RTK)
A second, optional engine. Where compression mode (above) shrinks the
conversation Slipstream sends to the model, RTK shrinks
the output of the shell commands your agent runs — git,
npm test, grep and 100+ others — before that output ever
reaches the model. It's bundled, so there's nothing to install.
Read why we added it →
-
Flip “Trim command output (RTK)”
Under Settings → Command compression. Slipstream wires RTK into Claude Code and puts the bundled
rtkon your path. It's off by default. -
Restart your agent
The hook loads when a Claude Code session starts, so restart any running sessions for it to take effect. Turning it off and restarting removes it cleanly.
RTK rewrites Bash commands only — Claude Code's built-in Read/Grep/Glob bypass it. This release wires Claude Code; other agents follow. On native Windows it falls back to a CLAUDE.md mode without the automatic rewrite (WSL recommended). And because RTK reports savings in tokens, the Commands figure on your dashboard is an estimate — it's a first implementation, and we're still gathering data on the real combined effect.
Semantic cache
When on, Slipstream answers repeated or near-identical requests locally instead of paying the model again — effectively free. Handy for agents that re-ask the same thing.
Daily budget
Set a dollar cap and Slipstream pauses upstream spend once you pass it for the day. Leave it empty for no limit. A safety net against a runaway agent loop.
Proxy
-
Port — where tools connect. Change it only if something else
already uses
8787. - Start proxy when Slipstream opens — savings begin the moment the app is running.
- Launch Slipstream at login — it waits quietly in the menu bar.
Models & pricing
The Models card holds each model's context window and price (in USD per million tokens, input and output). The context window guides how aggressively context is compressed; the prices drive the dollar estimates on your dashboard. Add a model or adjust one when a provider changes pricing — risky edits ask you to confirm against a diff first, so you can't fat-finger your numbers.
Privacy toggle
Share anonymous usage stats controls whether headroom sends anonymous aggregate figures (tokens saved, app version, OS — never your code or prompts) to its public dashboard. It's on by default upstream; one click turns it off. See Privacy & your data.
8 · Shared memory
Memory lets your agents remember facts, preferences and project context across sessions — and across each other. Something Claude Code learns can be recalled by Cursor later. It's stored locally and off by default.
-
Turn on memory
Open the Memory tab and click Turn on memory. Nothing is downloaded; it's all local.
-
Let your agents work
Memories accumulate as agents run. Each entry shows its text, which agent saved it, which project (workspace) it came from, when, and an importance score.
-
Search and prune
Use the search box to find a memory. If there's more than one memory space, a switcher lets you pick. To remove one, click the trash icon, then Forget? to confirm.
If memory was on and you turn it off, your existing memories stay readable — agents just stop saving new ones. Everything lives in your user directory and is yours to delete.
9 · MCP tools
Beyond the proxy, headroom ships an MCP server — a set of tools
your agent can call directly: compress, retrieve and
stats. Slipstream installs it for you. On the
Agents tab, find the MCP tools card and click
Install next to a client (such as Claude Desktop or Claude Code).
- Tools become available the next time that agent restarts.
- It works even while the proxy is off — the MCP server is independent.
- Click Remove to uninstall. If a client was set up outside Slipstream, manage it there.
10 · Privacy & your data
Slipstream is a local proxy on your own laptop. Requests go agent → localhost → your model provider. Nothing is routed through our servers, because there are no servers.
- No account, no sign-in, no cloud.
- Your agent keeps its own API keys and its own provider connection.
- Stats and memory live in a local database, yours to delete.
- Anonymous usage stats are surfaced honestly and are one click to disable.
- The compression core is open source, Apache 2.0.
To see exactly where things are kept, open Settings → Storage. Each row — Headroom data, Slipstream data, and the proxy log — has a Reveal button that opens the folder in your file manager.
11 · Your free month & plans
Slipstream is free for 30 days with everything included — no card, no account. The trial starts the moment you install. The menu-bar panel and Settings → License both show how many days are left.
After the trial, pick the plan that matches your AI spend. It's an honor system — every plan unlocks the same full app:
| If your AI plan is… | Slipstream |
|---|---|
| ~$20 / mo | $5 / mo |
| ~$100 / mo | $8 / mo |
| ~$200 / mo | $15 / mo |
-
Pick a plan
In Settings → License, choose the tier that fits and complete checkout in your browser.
-
Paste your key
Your license key (it starts with
SLIP-) arrives right after checkout. Paste it into the License field and click Activate.
Compression simply pauses — your agents keep working, just without the savings — and all your savings history stays intact. Subscribe whenever you're ready and it picks right back up.
12 · Updates
Slipstream updates itself. To check on demand, open Settings → About and click Check for updates. If one's available it downloads, verifies, installs, and relaunches the app. The same card shows your current app version and the bundled headroom version.
13 · Troubleshooting
The proxy won't start, or the dot is red
Usually the port is taken by another program. Open Settings → Proxy and change the Port to something else (for example 8788), then update any tools you connected by hand to the new address. If it still fails, quit and reopen Slipstream, then check the proxy log via Settings → Storage → Reveal.
I wrapped an agent but nothing is being saved
An agent reads its config only at startup. Restart the agent (or open a fresh session) after wrapping. Also confirm the proxy is running — the menu-bar dot should be green — and that the agent is the one you wrapped.
The dashboard is empty
Nothing has gone through yet. Make sure the proxy is running and at least one wrapped agent has made a request. If the proxy is running it will say it's listening and waiting; send a prompt from a wrapped agent and the numbers appear.
macOS won't let me open the app
Right-click the app → Open, or System Settings → Privacy & Security → Open Anyway. You only do this on first launch. See Install.
Does compression change my agent's answers?
The goal is the same answer from less input — headroom removes redundancy (repeated file reads, duplicate stack traces, boilerplate), not signal. The dashboard shows the per-request numbers so you can verify rather than trust. If a particular tool is sensitive, try Cache-friendly mode in Settings.
Will Slipstream see my API keys?
Your agent keeps its own keys and its own provider connection. Slipstream is a local pass-through proxy on your machine; requests go agent → localhost → provider. Nothing is routed through our servers, because there are none.
How do I uninstall?
First unwrap any connected agents on the Agents tab so their config is restored. Then remove the app the normal way for your OS (drag to Trash on macOS, delete the AppImage on Linux). To wipe data too, delete the folders shown in Settings → Storage.
14 · Reference
At a glance
| Default proxy URL | http://localhost:8787/v1 |
| Change the port | Settings → Proxy → Port |
| Supported agents | Claude Code, Codex, Cursor, Aider, GitHub Copilot |
| Other tools | Point the OpenAI base URL at the proxy |
| MCP tools | compress, retrieve, stats — Agents → MCP tools |
| Compression engine | headroom (Apache 2.0) |
| Command compression | RTK (Apache 2.0) |
| Where data lives | Settings → Storage → Reveal |
| Update | Settings → About → Check for updates |
Status colors
| Green | Proxy running — compressing requests. |
| Amber (pulsing) | Starting up or recovering. |
| Grey | Proxy stopped. |
| Red | Proxy failed to start — check the port and the log. |
Still stuck, or hit something this guide doesn't cover? The compression engine and its issues live on GitHub, and the homepage has a shorter FAQ.