All plugins
Plugin

RTK-AI Token Saver

Agents read a lot of command output. Every git status, git log, directory listing and test run lands in their context. This plugin routes an agent’s shell commands through RTK, which keeps the facts and drops the noise before the model reads them.

Built on RTK, open source by rtk-ai under the Apache-2.0 license:github.com/rtk-ai/rtk

Download for WindowsSee it work

−92%

less git status output in our test run

−25%

across six everyday commands in the same run

0

restarts to switch it on or off

On the canvas

Arrow on, and the next command is compressed

The same git status from the same agent. Real output of git and of RTK 0.50.0 in a small repository.

FrontendPricing page
> What have I changed so far?
●Bash(git status)
⎿
On branch main
Changes to be committed:
  (use "git restore --staged <file>..." to unstage)
        modified:   src/pricing/Toggle.tsx

Changes not staged for commit:
  (use "git add <file>..." to update what will be committed)
  (use "git restore <file>..." to discard changes in working directory)
        modified:   src/api/checkout.ts
        modified:   src/pricing/Plans.tsx

Untracked files:
  (use "git add <file>..." to include in what will be committed)
        src/pricing/Badge.tsx
        src/pricing/plans.test.tsx
●Three files modified (Toggle.tsx is staged), two new files in src/pricing.
RTK-AI Token Saver

Connect an agent

Draw an arrow from an agent to this card: its git, ls, grep and test commands come back compressed, starting with the next one.

What it does

Same facts, fewer tokens

RTK is a command-line proxy. It runs the real tool and prints a compact version of what it said.

Your agent’s shell commands, through RTK

git, ls, grep, find, test runners, linters: when RTK knows a command, git status becomes rtk git status. When it doesn’t, the command runs as it is.

Compressed before the model reads it

RTK groups, trims and de-duplicates the output. A long status becomes a short list of changed files, and a passing test run can shrink to a line.

Only shell output

Your prompts, the agent’s answers and its own file-reading tools are untouched. An agent that needs the full output of one command can put RTK_DISABLED=1 in front of it.

Measured

What it saved in one real test

Claude Code on Windows, RTK 0.50.0, the plugin’s card reading RTK’s own numbers. The more boilerplate a command prints, the more it saves.

  • git status−92%
  • ls -la src/core−62%
  • git log -n 20−32%
  • git log -n 3−18%
  • git log -n 10 --stat0%
  • grep · small result0%

All six commands

≈ 794 of ≈ 3.19k tokens

−25%

One test run on 2026-09-29, not a benchmark. Results depend on your repository and your commands. RTK estimates tokens as bytes ÷ 4, so the percentages are exact and the token counts approximate.

How it works

Nothing to configure

The card installs RTK, and the arrow does the rest.

  1. Add the card

    Add-card menu → “Plugin…” → Plugins tab → RTK-AI Token Saver → “Add to canvas”.

  2. Get RTK

    Press “Download RTK”: the official build comes from GitHub (about 4.5 MB) and must match the SHA-256 checksum GitHub publishes, or it isn’t used. An rtk you’ve already installed is found on its own.

  3. Draw an arrow

    From an agent to the card. Its next shell command already goes through RTK. No restart.

  4. Remove it to stop

    Take the arrow away and the next command runs raw again. The card keeps count of what was saved, per agent.

Permissions

Never an extra prompt

RTK never adds a permission prompt you wouldn’t have had, and never approves a command that would have asked.

  • Dangerous mode

    Every command RTK knows is compressed.

  • Normal mode — Claude Code

    Read-only commands like git status, git log, ls, cat and grep are compressed with no prompt, just as they ran before. Commands that would ask anyway still ask once, for the rtk form. Everything else runs unchanged.

  • Normal mode — Cursor, OpenCode, Kilo Code

    Their permission prompts can’t be predicted from outside, so RTK works for them in dangerous mode only.

  • Canvas mode

    The agent has no shell of its own there, so there’s nothing to compress.

Agent CLIs

Works with four CLIs today

The ones that let a command be rewritten right before it runs.

Claude Code

Normal & dangerous modeAgent guide

Cursor CLI

Dangerous modeAgent guide

OpenCode

Dangerous modeAgent guide

Kilo Code

Dangerous modeAgent guide

Other agents can be connected, but the card marks them “Not supported”: their CLIs don’t offer a way to rewrite a command before it runs yet.

Safe by default

If anything goes wrong, you get the raw command

The plugin only ever swaps a command for RTK’s version of the same command.

Falls back on any failure

RTK missing, crashing, too slow, or not knowing a command: the original command runs unchanged.

Checked download

The RTK build is only used if its SHA-256 checksum matches the one GitHub publishes. No checksum, no download.

Telemetry off

RTK’s own telemetry is opt-in, and NeuroSquad switches it off for every agent it starts, whatever it’s set to.

Your settings untouched

Your RTK config and your CLI’s settings stay as they are. RTK’s history is kept per agent in NeuroSquad’s data folder.

Credit

RTK is made by rtk-ai

The compression is RTK’s work, published as open source under the Apache-2.0 license. NeuroSquad downloads the official build unchanged, or uses the one you installed, and decides when an agent’s command goes through it.

rtk-ai/rtk on GitHubApache-2.0
FAQ

Questions about RTK-AI Token Saver

Fewer tokens from the next command

Download NeuroSquad, add RTK-AI Token Saver, and draw an arrow.

Download for WindowsRead the guide
Windows 10 / 11 · 64-bitAll plugins