Open Source · MIT · Rust 50.3k stars

Your AI agent is drowning
in CLI noise. Fix it.

rtk compresses command outputs before they reach the context window. Better reasoning. Longer sessions. Lower costs.

89% avg. noise removed
3x longer sessions
30+ commands
Why RTK? The numbers.

89% noise reduction measured across 2,900+ real-world dev commands: cargo test 91.8%, git status 80.8%, find 78.3%, grep 49.5%. Free, open source (MIT, Rust).

Install rtk Star on GitHub
Viking warrior smashing token noise with his axe
Terminal
$ rtk gain

📊 RTK Token Savings
════════════════════════════════════════

Total commands:    2,927
Input tokens:      11.6M
Output tokens:     1.4M
Tokens saved:      10.3M (89.2%)

By Command:
────────────────────────────────────────
Command               Count      Saved     Avg%
rtk find                324       6.8M    78.3%
rtk git status          215       1.4M    80.8%
rtk grep                227     786.7K    49.5%
rtk cargo test           16      50.1K    91.8%

$
Works with
  • Claude Code
  • Cursor
  • Aider
  • Gemini CLI
  • OpenAI Codex
  • Cline
  • Windsurf
  • GitHub Copilot
// context forge

One forge. Three tools.

RTK, ICM, and Vox share one philosophy: open source, Rust, zero telemetry, local-first.

RTK Compress

Commands, file reads, tests — compressed 60-90% before they reach your model context. Zero config.

89% noise removed
ICM Remember

Persistent memory across sessions. Your agent picks up where it left off — decisions, errors, context.

Infinite context memory
Vox Speak

Voice output for your AI agent. Three TTS backends, four Claude Code integration modes.

3 TTS backends
// 01 — the problem
Viking bracing against a tidal wave of tokens

The problem with AI coding today

Every command your agent runs pollutes the context window with noise. Here's what that costs you.

Context pollution

Your 200K context window isn't infinite. When cargo test dumps 5,000 tokens of boilerplate, that's 5,000 tokens less for reasoning about your actual code.

context_quality: degraded ▼

Sessions too short

Context overflows, the agent restarts, you lose the thread. On flat-rate plans, you hit rate limits 40% faster than you should.

session_remaining: 32% ▼

Costs that explode

On pay-per-token setups (API, Gemini CLI, Aider), 70% of your bill is noise the LLM doesn't need. A team of 10 wastes ~$1,750/month.

token_waste: $1,750/mo ▲
Viking compressing tokens into a glowing crystal

See the difference

Real outputs, real savings. Side-by-side comparison on actual commands.

cargo test ~4,823 tokens 
warning: unused variable: `start`
   --> src/init.rs:561:17
    |
561 |             let start = i;
    |                 ^^^^^ help: prefix it with an underscore

warning: `rtk` (bin "rtk" test) generated 17 warnings
    Finished `test` profile target(s) in 0.20s
     Running unittests src/main.rs

running 262 tests
test cargo_cmd::tests::test_filter_cargo_build_success ... ok
test cargo_cmd::tests::test_filter_cargo_clippy_clean ... ok
test cargo_cmd::tests::test_filter_cargo_test_all_pass ... ok
test cargo_cmd::tests::test_filter_cargo_test_failures ... ok
test cc_economics::tests::test_compute_dual_metrics ... ok
test git::tests::test_compact_diff ... ok
test git::tests::test_filter_branch_output ... ok
test git::tests::test_filter_log_output ... ok
test git::tests::test_filter_status_with_args ... ok
... 253 more tests, all passing

test result: ok. 262 passed; 0 failed; 0 ignored; 0 measured
rtk cargo test ~11 tokens -99% 
✓ cargo test: 262 passed (1 suite, 0.08s)
pytest -v ~756 tokens 
===== test session starts ======
platform darwin -- Python 3.14.3, pytest-9.0.2
collected 33 items

test_utils.py::TestStringUtils::test_strip PASSED    [  3%]
test_utils.py::TestStringUtils::test_upper PASSED    [  6%]
test_utils.py::TestStringUtils::test_split PASSED    [  9%]
test_utils.py::TestMathUtils::test_sqrt PASSED       [ 27%]
test_utils.py::TestMathUtils::test_ceil PASSED       [ 30%]
test_utils.py::TestJsonUtils::test_dumps PASSED      [ 51%]
test_utils.py::TestJsonUtils::test_loads PASSED      [ 54%]
test_utils.py::TestPathUtils::test_join PASSED       [ 63%]
test_utils.py::TestListOps::test_sort PASSED         [ 78%]
test_utils.py::TestListOps::test_flatten PASSED      [100%]
... 23 more tests, all passing

====== 33 passed in 0.05s ======
rtk test "pytest -v" ~24 tokens -96% 
33 passed in 0.02s
go test ./... -v ~592 tokens 
=== RUN   TestControllers
Running Suite: Controller Suite
Random Seed: 1771093287
Will run 1 of 1 specs
•
Ran 1 of 1 Specs in 6.037 seconds
SUCCESS! -- 1 Passed | 0 Failed | 0 Pending
--- PASS: TestControllers (6.04s)
PASS
ok  kubecraft.ai/.../controller  6.610s
=== RUN   TestNewAWSBedrockClient
--- PASS: TestNewAWSBedrockClient (0.00s)
=== RUN   TestNewAzureOpenAIClient
--- PASS: TestNewAzureOpenAIClient (0.00s)
--- PASS: TestAzureOpenAIGenerateCode_Success (0.00s)
--- PASS: TestNewMistralClient (0.00s)
--- PASS: TestGenerateCode_Success (0.00s)
PASS
ok  kubecraft.ai/.../llm  0.776s
rtk test "go test ./... -v" ~246 tokens -58% 
--- PASS: TestControllers (5.07s)
ok  kubecraft.ai/.../controller  5.847s
--- PASS: TestNewAWSBedrockClient (0.00s)
--- PASS: TestNewAzureOpenAIClient (0.00s)
--- PASS: TestAzureOpenAIGenerateCode_Success (0.00s)
--- PASS: TestNewMistralClient (0.00s)
--- PASS: TestGenerateCode_Success (0.00s)
ok  kubecraft.ai/.../llm  0.246s
git diff HEAD~1 ~21,500 tokens 
diff --git a/index.html b/index.html
index 1b7488b..0ebac4f 100644
--- a/index.html
+++ b/index.html
@@ -629,7 +629,7 @@
       width: 100%;
       border-collapse: collapse;
       font-size: 0.88rem;
-      min-width: 800px;
+      min-width: 1050px;
     }
@@ -1051,6 +1051,114 @@
+    /* === Share My Gain === */
+    .share-gain { background: var(--bg); }
+    .share-gain-card {
+      max-width: 600px;
+      margin: 0 auto;
... 855 insertions(+), 266 deletions(-)
rtk git diff HEAD~1 ~1,259 tokens -94% 
index.html | 1121 ++++++++++++------
 1 file changed, 855 ins(+), 266 del(-)

--- Changes ---

index.html
  @@ -629,7 +629,7 @@
  -      min-width: 800px;
  +      min-width: 1050px;
  @@ -1051,6 +1051,114 @@
  +    /* === Share My Gain === */
  +    .share-gain { background: var(--bg); }
  ... (truncated)
git status ~120 tokens 
On branch master
Your branch is up to date with 'origin/master'.

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:   index.html
	modified:   src/main.rs
	modified:   src/config.rs

Untracked files:
  (use "git add <file>..." to include in what will be committed)
	.fastembed_cache/
	tests/

no changes added to commit
rtk git status ~30 tokens -75% 
* master...origin/master
~ 3  index.html, src/main.rs, src/config.rs
? 2  .fastembed_cache/, tests/
git log --stat -10 ~1,430 tokens 
commit c84fa3c fix: add website URL (rtk-ai.app)
Author: patrick szymkowiak <52030887+pszymkowiak@...>
Date:   Thu Feb 12 12:28:37 2026 +0100

    - Cargo.toml: add homepage field
    - README.md: add website/github/install links

 .github/workflows/release.yml | 4 ++--
 Cargo.toml                    | 1 +
 Formula/rtk.rb                | 2 +-
 README.md                     | 8 ++++++++
 5 files changed, 14 insertions(+), 3 deletions(-)

commit a0d2184 feat(ci): automate Homebrew formula
... 8 more commits with full stats
rtk git log -n 10 ~194 tokens -86% 
c84fa3c fix: add website URL (rtk-ai.app) (#81)
a0d2184 feat(ci): automate Homebrew formula (#80)
55d010a fix: update stale repo URLs (#78)
9e764c4 chore(master): release 0.13.1
bd76b36 fix(ci): fix release workflow artifacts
0d388b3 chore(master): release 0.13.0
93364b5 feat(sqlite): custom db location
3dde699 chore(master): release 0.12.0
645a773 feat(cargo): cargo install filtering
22182e4 chore(master): release 0.11.0
cat src/main.rs ~10,176 tokens 
mod cargo_cmd;
mod cc_economics;
mod ccusage;
mod config;
mod container;
mod curl_cmd;
mod deps;
mod diff_cmd;
mod discover;
mod filter;
mod find_cmd;
mod gain;
mod gh_cmd;
mod git;
mod grep_cmd;
mod init;
mod json_cmd;
mod ls;
mod main.rs
... 1,295 lines total (enums, match arms, CLI parsing)
rtk read src/main.rs -l aggressive ~504 tokens -95% 
use anyhow::{Context, Result};
use clap::{Parser, Subcommand};
struct Cli {
    // ... implementation
enum Commands {
    // ... implementation
enum GitCommands {
    // ... implementation
enum CargoCommands {
    // ... implementation
fn main() -> Result<()> {
    // ... implementation
grep -rn "pub fn" src/ ~2,108 tokens 
src/cargo_cmd.rs:17:pub fn run(cmd: CargoCommand, ...) -> Result<()>
src/cargo_cmd.rs:551:pub fn run_passthrough(args: ...) -> Result<()>
src/cc_economics.rs:184:pub fn run(...)
src/config.rs:73:    pub fn load() -> Result<Self>
src/config.rs:85:    pub fn save(&self) -> Result<()>
src/config.rs:97:    pub fn create_default() -> Result<PathBuf>
src/filter.rs:57:    pub fn from_extension(ext: &str) -> Self
src/filter.rs:309:pub fn get_filter(level: FilterLevel) -> Box<dyn FilterStrategy>
... 112 matches across 47 files
rtk grep "pub fn" src/ ~940 tokens -55% 
112 in 47 files:

src/cargo_cmd.rs (3):
    17: pub fn run(cmd: CargoCommand, ...) -> Result<()>
   551: pub fn run_passthrough(args: ...) -> Result<()>

src/config.rs (4):
    73: pub fn load() -> Result<Self>
    85: pub fn save(&self) -> Result<()>
    97: pub fn create_default() -> Result<PathBuf>
   109: pub fn show_config() -> Result<()>

src/git.rs (2):
    22: pub fn run(cmd: GitCommand, ...) -> Result<()>
  1264: pub fn run_passthrough(args: ...) -> Result<()>
... +62
find . -name "*.rs" ~276 tokens 
./target/debug/build/serde_core-.../out/private.rs
./target/debug/build/libsqlite3-sys-.../out/bindgen.rs
./src/ls.rs
./src/local_llm.rs
./src/learn/detector.rs
./src/learn/report.rs
./src/discover/registry.rs
./src/discover/provider.rs
./src/cargo_cmd.rs
./src/ccusage.rs
./src/config.rs
./src/find_cmd.rs
./src/gain.rs
./src/git.rs
... 49 files total
rtk find "*.rs" . ~149 tokens -46% 
49 files in 4 dirs:

src/ cargo_cmd.rs cc_economics.rs
  ccusage.rs config.rs container.rs
  curl_cmd.rs deps.rs diff_cmd.rs
  filter.rs find_cmd.rs gain.rs
  gh_cmd.rs git.rs grep_cmd.rs
  init.rs json_cmd.rs ls.rs
  main.rs ...
src/discover/ mod.rs provider.rs
  registry.rs report.rs
src/learn/ detector.rs mod.rs report.rs
src/parser/ error.rs formatter.rs
ls -la src/ ~3,200 tokens 
total 928
drwxr-xr-x  41 patrick  staff   1312 .
drwxr-xr-x  25 patrick  staff    800 ..
-rw-r--r--   1 patrick  staff  16394 cargo_cmd.rs
-rw-r--r--   1 patrick  staff  27220 cc_economics.rs
-rw-r--r--   1 patrick  staff   9503 ccusage.rs
-rw-r--r--   1 patrick  staff   2884 config.rs
-rw-r--r--   1 patrick  staff  12886 container.rs
-rw-r--r--   1 patrick  staff   3406 curl_cmd.rs
-rw-r--r--   1 patrick  staff  12156 filter.rs
-rw-r--r--   1 patrick  staff  10648 gain.rs
... 20 more lines
rtk ls src/ ~640 tokens -80% 
discover/
parser/
cargo_cmd.rs  16.0K
cc_economics.rs  26.6K
ccusage.rs  9.3K
config.rs  2.8K
container.rs  12.6K
filter.rs  11.9K
find_cmd.rs  2.9K
gain.rs  10.4K
...

37 files, 2 dirs (37 .rs)
cat Cargo.toml ~368 tokens 
[package]
name = "rtk"
version = "0.13.1"
edition = "2021"
license = "MIT"
homepage = "https://www.rtk-ai.app"
repository = "https://github.com/rtk-ai/rtk"

[dependencies]
clap = { version = "4", features = ["derive"] }
anyhow = "1.0"
ignore = "0.4"
walkdir = "2"
regex = "1"
lazy_static = "1.4"
serde = { version = "1", features = ["derive"] }
serde_json = { version = "1" }
colored = "2"
dirs = "5"
rusqlite = { version = "0.31", features = ["bundled"] }
... + cargo-deb, cargo-generate-rpm metadata
rtk deps ~55 tokens -85% 
Rust (Cargo.toml):
  Dependencies (15):
    clap (4)
    anyhow (1.0)
    ignore (0.4)
    walkdir (2)
    regex (1)
    lazy_static (1.4)
    serde (1)
    serde_json (1)
    colored (2)
    dirs (5)
    ... +5 more
Viking king on throne with 88.9% holographic display

Real-world savings

Actual rtk gain output from a happy developer.

A developer's feedback

After a few weeks of daily use: 15,720 commands processed, 138M tokens saved. Just run rtk gain to see yours.

88.9% efficiency
rtk gain dashboard showing 88.9% efficiency

Detailed breakdown

Daily, weekly, and monthly stats by command. Track your savings over time.

Per-command analytics
rtk gain daily and weekly breakdown
Viking squad marching through datacenter corridor

No AI tool offers unlimited usage. Every token counts.

Even at $200/mo, every tool has caps. RTK compresses CLI noise so your limits stretch further.

A typical 2h coding session with an AI agent:

~60 CLI commands run
~210K tokens of CLI noise
~23K with RTK (89% less)

Without RTK, CLI output alone can overflow a 200K context window. Based on avg 3,500 tokens/command measured across 2,900+ real commands.

Claude Code Terminal
Price$20 — $200/mo
Limits~45 msgs/5h (Pro), 5-20x on Max
Context200K tokens
Sessions ~3x longer
Even Max $200/mo (20x Pro) has weekly caps (240-480h). Quota resets every 5h. RTK compresses CLI outputs by avg 89%, so each message carries less noise and your quota stretches ~3x.
Cursor IDE
Price$20 — $200/mo
Limits$20 credits/mo (Pro), ~225 Claude reqs
ContextUp to 200K (Max mode)
Credits go ~2x further
Even Ultra $200/mo (20x credits) is capped. Each request consumes credits based on model — Claude burns 2.4x faster than Gemini. RTK compresses CLI outputs so each request starts cleaner.
OpenAI Codex Agent
Price$20/mo (Plus) — $200/mo (Pro)
Limits30-1,500 msgs/5h by plan
Context192K tokens
More iterations per cap
Included with ChatGPT plans. Pro $200/mo caps at 1,500 msgs/5h. The agent runs commands autonomously — each output eats your cap. RTK compresses them for more iterations per window.
Windsurf IDE
Price$15 — $60/mo
Limits500 credits/mo (Pro)
Context200K tokens
Credits last ~2x longer
Enterprise $60/user gets 1,000 credits/mo — still capped. Cascade consumes credits per prompt. RTK compresses CLI outputs so each interaction uses fewer tokens, stretching your credits.
Gemini CLI Terminal
PriceFree — pay-per-token
Limits1,000 req/day, 60 req/min (free)
Context1M tokens
~70% less on token bill
Free tier is generous (1,000 req/day) but still rate-limited. Beyond that, you pay per token. RTK compresses CLI outputs by avg 89%, cutting your bill or freeing rate limit headroom.
Aider Terminal
PriceFree + API costs ($5-300+/mo)
LimitsPer API provider
ContextPer model (up to 200K)
~70% less API cost
BYO API key — you pay per token to OpenAI, Anthropic, etc. RTK compresses every command output before it reaches the model, directly cutting your API bill by ~70% on CLI-heavy workflows.
GitHub Copilot IDE
PriceFree — $39/mo (Pro+)
Limits50-1,500 premium req/mo
ContextPer model (up to 200K)
Better context quality
Enterprise $39/user: 1,000 premium req/mo. Base completions are unlimited, but Chat and the coding agent have caps. RTK keeps terminal output lean so premium requests carry more useful context.
Cline / Roo VS Code
PriceFree + API costs ($0-500+/mo)
LimitsPer API provider
ContextPer model (up to 200K)
~70% less API cost
No tool-side limit, but your API provider caps apply. Heavy users report $200-500+/mo. RTK compresses every output by avg 89%, directly cutting your bill and reducing context overflow.

Pricing verified Feb 2026. Limits vary by usage and plan. RTK savings based on avg 89% compression across 2,900+ real commands.

Coming Soon

RTK Cloud

Visibility and control over your team's AI coding costs. Know what's wasted. Fix it.

Token analytics

Dashboard per dev, per project, per tool

Team savings reports

"Your team saved $4,200 this month"

Rate limit alerts

Monitoring & smart notifications

Enterprise controls

SSO, audit logs, compliance

Free for open-source. Teams from $15/dev/month.

0 teams on the waitlist

No spam. One email when we launch.

Viking slamming axe into ground creating green energy shockwave
// get started

Get started in 30 seconds

Install, activate the auto-rewrite hook, and every command is compressed automatically.

Quick Install

One-liner for Linux & macOS

curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/refs/heads/master/install.sh | sh

Via Homebrew

macOS & Linux

brew install rtk
brew upgrade rtk

Pre-built Binaries

macOS, Linux, Windows

Download from Releases →

Then activate the auto-rewrite hook

rtk init --claude-code

Installs PreToolUse hook in Claude Code settings.json — every Bash call is rewritten automatically. Full install guide →

1 curl ... | sh
2 rtk init --claude-code
3 rtk gain
Viking raising axe in triumphant war cry
// ship it

Your AI doesn't need
to read all that.

Install rtk. Better code, longer sessions, lower costs.

Install rtk Star on GitHub

Frequently asked questions

What is RTK (Rust Token Killer)?

RTK is an open-source CLI tool that compresses command outputs before they reach the AI context window. It reduces token usage by 60-90% with zero configuration changes, enabling longer AI coding sessions and lower API costs. RTK is written in Rust, MIT licensed, and works transparently with Claude Code, Cursor, and any terminal-based AI assistant.

How many tokens does RTK actually save?

Based on measurements across 2,900+ real-world commands, RTK removes an average of 89% of CLI output noise. Command-level savings: cargo test (91.8% savings), git status (80.8%), find (78.3%), grep (49.5%). A developer who ran 15,720 commands saved 138 million tokens over several weeks, tracked live via rtk gain.

Is RTK free? Are there usage limits?

RTK is completely free. It is open source under the MIT license, with source code available on GitHub at github.com/rtk-ai/rtk. There are no usage limits, no API keys required, no telemetry, and no accounts. RTK Cloud (waitlist) will offer additional team features.

Which AI coding tools and commands does RTK support?

RTK works with Claude Code (Anthropic), Cursor, Gemini CLI, Aider, and any AI assistant that reads terminal output. It supports all major CLI commands: cargo test, pytest, go test, git diff, git status, git log, grep, find, ls, pnpm list, tsc, eslint, prisma, docker, kubectl, and more.

How does RTK work without changing my workflow?

Running rtk init --global installs a PreToolUse hook in Claude Code that automatically rewrites Bash commands to rtk equivalents at the proxy layer. You continue using your normal commands. RTK intercepts and compresses the output before it enters the context window, with no manual changes to prompts or workflows needed.

Does RTK affect code quality or AI reasoning accuracy?

No. RTK removes verbose boilerplate and repetitive output noise, not meaningful content. Test failures, error messages, diffs, and stack traces are preserved in full. The AI receives the same essential information with 89% less noise, which typically improves reasoning quality by reducing context pollution.