densilo
New Measured on real Claude Code and Codex sessions

More work before you hit your usage limit.
Up to 53% fewer tokens on Claude Code, 57% on Codex.

For Claude Code and Codex, on your machine or in your pull requests. Your prompts and code never reach our servers.

On your machine Claude Code and Codex: up to 53% and 57% fewer tokens on long sessions Install in one command For teams on GitHub Cut your Claude PR review cost in half, and get twice as many reviews finished Add Densilo to your Claude action

Available for macOS (Apple silicon), Linux (x64) and Windows (x64, beta).

Works with the coding agent you already use

Claude Code CLI Codex CLI beta Editors (VS Code, Cursor), OpenCode, Aider: planned

Measured on real sessions

The longer the session, the more it saves.

About 75% of what Claude Code costs is re-reading context it already sent. Densilo cuts that down. We replayed real recorded Claude Code and Codex sessions through it and counted what the model would receive. Short sessions change little; savings grow with session length. How we measured →

up to53%fewer tokens on long Claude Code sessions

The best among Claude Code sessions long enough for Densilo to work on; the median was 36%. Across all Claude Code sessions: 39% fewer tokens, 34% lower input cost.

Claude Code results
up to57%fewer tokens on long Codex sessions

The best among Codex sessions long enough for Densilo to work on; the median was 32%. Across all Codex sessions: 31% fewer tokens, 27% lower input cost.

Codex results
0 uploadsto see your own number

Run densilo benchmark in a repository. It replays your own recent sessions on your machine, offline. No model is called and nothing is sent to us.

Measure yours

How it works

Install once, then code the way you already do.

Densilo runs on your machine, between your coding agent and the model. It sends fewer tokens every turn, and your agent keeps the context it needs. See your own number, offline:

densilo benchmark

  1. Install Densilo

    One command downloads the app, verifies its checksum and adds the densilo command. On macOS and Linux, open a new terminal window afterwards so it finds densilo; on Windows it works in the same window.

    macOS and Linux

    curl -fsSL https://api.densilo.com/install.sh | sh

    Windows (x64) — beta, in PowerShell

    irm https://api.densilo.com/install.ps1 | iex

  2. Sign in

    Run this in the new terminal window. Your browser opens to sign in with Google (on a remote machine, open the link it prints on any device). Until you sign in, Densilo stays out of the way and your agent runs as it always has. If the terminal says densilo isn't found, run ~/.local/bin/densilo login instead; on Linux, logging out and back in once makes the plain command work everywhere.

    densilo login

  3. Start your coding tool through Densilo

    Use it exactly as before; Densilo sends fewer tokens on your behalf. Start Claude Code with densilo run claude instead of claude: a plain claude runs without Densilo. In the terminal, Densilo runs only for that session and writes nothing to your settings.

    Claude Code in the terminal

    densilo run claude

    Codex CLI (beta)

    densilo run codex

  4. Watch the savings

    On a Mac, the menu bar shows your running sessions and today's savings. The dashboard shows every day's totals.

    densilo status

Correctness & privacy

Built to be trusted with your codebase.

"I can't send our code to another company."

Your prompts stay on your machine

Densilo runs inside the app on your machine. We never receive your prompts, code or model responses. Only usage counts, such as token totals and the model name, are sent, to show your usage.

"Fewer tokens means the model misses something."

Your agent keeps the context it needs

Densilo cuts what your agent resends, not what it needs to do the work. Try it on your own sessions and judge the results for yourself.

"One more tool in the loop is one more thing to break."

Nothing to break

Densilo lives only as long as the session you start. If it ever can't start, your agent runs directly against the model instead.

Download

Get Densilo for your computer.

macOS

Available

Apple Silicon (M1 and later) · macOS 11+ · version 0.3.14

curl -fsSL https://api.densilo.com/install.sh | sh

Paste into Terminal. This is the recommended install: it also adds the densilo command.

Download .zip instead

Move Densilo to Applications and open it; it adds the densilo command on first launch. If macOS blocks the app, go to System Settings → Privacy & Security → Open Anyway.

Intel Mac? Coming soon.

Windows

Beta

Windows (x64) — beta · Windows 10 and 11 · version 0.3.14

irm https://api.densilo.com/install.ps1 | iex

Paste into PowerShell. Installs to %LOCALAPPDATA%\Densilo and adds the densilo command to your PATH; no admin rights or Python needed. Then run densilo login and densilo run claude.

Download .zip instead

The beta is not code-signed yet. If Windows warns that the app is from an unknown publisher, choose More info → Run anyway.

Linux

Available

x86_64 · glibc 2.28+ (Ubuntu 20.04+, Debian 11+, RHEL/Rocky 8+, Fedora, Amazon Linux 2023) · version 0.3.14

curl -fsSL https://api.densilo.com/install.sh | sh

Paste into a terminal. Installs the densilo command to ~/.local/bin (or /usr/local/bin as root). No Python needed. There's no menu-bar icon on Linux; use densilo status and the dashboard. Open a new terminal after installing; on some desktops, log out and back in once so every terminal finds densilo.

Download .tar.gz instead

arm64 Linux? Coming soon.

Running Claude on pull requests? Cut the review cost in half and get twice as many reviews finished. Densilo for teams →

Questions

Frequently asked questions

How much will I save?

It depends on how long your sessions run. In replays of real recorded sessions, Claude Code sessions long enough for Densilo to work on sent a median 36% fewer tokens (up to 53%), and Codex sessions a median 32% fewer (up to 57%). On a Claude Pro or Max plan, that's 2× more work before you hit your usage limit. Short sessions change little. Run densilo benchmark to see your own number, offline. See the results.

Does it change what my agent can do?

Your agent keeps the context it needs to do the work. A replay measures tokens and cost, not task quality, so the best test is your own: use Densilo on real work and compare.

What do you see of my code or prompts?

None of it. Densilo runs on your machine. The app sends us only usage counts (tokens, requests and the model name), plus the app version and your operating system, so your dashboard can show usage and savings. Privacy policy.

What happens if Densilo stops working?

Your agent keeps working. If Densilo can't start, or your plan has run out, requests go straight to the model, as they would without Densilo. Densilo never blocks your agent.

Which coding tools does it work with?

Claude Code in the terminal, Codex CLI (beta), and Claude reviews in GitHub Actions. Editor extensions (VS Code, Cursor, Windsurf), OpenCode and Aider are planned and not yet tested.

See your own number, then decide.

Install, run densilo benchmark in a repository you work in, and see what Densilo would have saved.

Download Densilo See the results