Same work.
60% fewer tokens.
Terminal AI coding agent. Every prompt trimmed. Every reply terse. Every output compressed.
MIT Licensev1.1.3
Same work.
Real numbers
This data is based on 1 prompt to build this site.
517.5K
tokens saved
39,710
context used
425
optimizations applied
What it doesSame coding agent.
Same coding agent.
Leaner every step.
BJIR wraps your existing CLI coding agent and applies aggressive token optimization at every layer: input, context, output, and tool I/O.
It runs locally. No proxy, no cloud, no data leaves your machine. Every compression technique is compiled in and runs in-process.
The result: same quality work, dramatically lower cost, faster responses, and sessions that last 2–3x longer before hitting limits.
Optimizations
Six layers of token compression.
Each optimization targets a different part of the token pipeline. Together they compound.
context management
33 ops249.0K saved
Context Prune
Drops stale conversation history each turn. Keeps what matters, discards what doesn't. The single biggest token saver, 249K tokens in one session.
web fetching
105 ops124.1K saved
I/O: Webfetch
Compresses web fetch results before they enter context. Strips HTML noise, collapses whitespace, extracts only relevant content from URLs.
output compression
114 ops15.1K saved
SmartCrush + Response Control
Crushes JSON payloads to minimal form. Compresses model output: abbreviates prose, strips filler, enforces terse responses. 114 operations in a single session.
shell & process wrapping
99 ops7.9K saved
I/O: Bash + RTK
Intercepts bash command output and wraps shell processes (git, cargo, npm, docker). Keeps only actionable lines: errors, warnings, changed files. Drops repetitive noise.
input trimming
39 ops10 saved
Prompt Refine
Trims prompts before they reach the model. Strips redundant whitespace, collapses repeated context, compresses system instructions.
read caching
35 ops18 saved
I/O: Read Dedup
Caches file reads and skips re-sending content already seen this session. Small per-read savings that compound across large codebases.
Works with
ClaudeOpenAIGeminiLlamaMistralGroqOllamaTogetherDeepSeekOpenRouterLM StudioClaudeOpenAIGeminiLlamaMistralGroqOllamaTogetherDeepSeekOpenRouterLM StudioClaudeOpenAIGeminiLlamaMistralGroqOllamaTogetherDeepSeekOpenRouterLM StudioClaudeOpenAIGeminiLlamaMistralGroqOllamaTogetherDeepSeekOpenRouterLM Studio
Get started
Install in 10 seconds.
1
Install the package
One global command. Works on macOS, Linux, and WSL.
2
Launch BJIR
Run bjir in any project directory. Connects to your configured provider.
3
Start coding
Same agent, same capabilities. Just fewer tokens per request.
4
Check your savings
Run bjir gain to see token reduction metrics for your session.
Install from source
$ git clone https://github.com/gogetrekt/bjir.git
$ cd bjir
$ cargo install --path .
CLI Reference
Commands.
Every command runs locally. No telemetry, no cloud calls.