Cut LLM Token Costs by 90%

A pluggable CLI filter that strips verbose output before it hits Claude, ChatGPT, or any LLM API. Save hundreds per month on token costs.

Get Started on GitHub

Real CLI output filtering in action

The Problem: Burning Money on Verbose Output

Every time you pipe a CLI command to Claude or ChatGPT, you pay for every character. When you run kubectl get -o yaml, you get 10,000 lines of output. Your LLM needs about 50 lines to make a decision. You're paying 200x for noise.

Error traces with 500 lines of stack unwinding. Git logs with 1000 commits. Build logs with duplicated status lines. Deploy manifests with hundreds of unchanged fields. All of it hits your API, all of it costs tokens, and 95% of it never influences the LLM's answer.

❌ Without Lowfat

kubectl get pods output

1,847 tokens

Cost: $0.18 (@ $0.10 per 1M input)

Or: ~$20 per month for 100 queries

✓ With Lowfat

Filtered output

156 tokens

Cost: $0.02 (91.6% savings)

Or: ~$2 per month for 100 queries

How Lowfat Works

Lowfat sits between your CLI and your LLM. It understands command patterns and strips the noise: timestamps, repeated lines, empty fields, verbose headers, debug metadata. What's left is signal.

One single binary. Works as an agent hook. Works as a shell wrapper. Works with custom filter plugins. Drop it in and it just works.

$ lowfat history --all
lowfat plugin candidates
─────────────────────────────────────────────────────────
  #  command                runs   avg raw   cost   savings   source
  1  kubectl get pods        24     1847t    $0.18  91.6%     kubectl
  2  git log --oneline       18      892t    $0.09  84.2%     git
  3  docker ps -a            15      634t    $0.06  79.1%     docker
      

Two months of real-world use: 91.8% token reduction across common DevOps and deployment workflows. Compounded across a team, that's hundreds of dollars per month saved.

Features

Single Binary, Zero Dependencies

Lowfat is a Go binary that compiles to a single executable. Drop it in your PATH, add it to your agent hooks, and it works. No runtime, no dependencies, no config hell.

Works Anywhere

Use it as an agent hook, a shell wrapper, a piped filter, or integrated into your scripts. Works with Claude, ChatGPT, Anthropic SDK, or any LLM API that accepts text input.

Plugin System for Custom Filters

Out of the box, Lowfat knows kubectl, docker, git, npm, and common Unix tools. Add your own filters via a simple plugin system. Customize per command.

Cost Tracking Built In

lowfat history --all shows exactly what you're saving. Track cost savings over time, per command, per team member. Know your ROI.

Open Source & Developer-Native

Runs locally. Runs offline. Open source licensing. No cloud lock-in. No telemetry. No vendor dependencies. Your data stays yours.

Real Results

Not theoretical. 91.8% token reduction verified over 2 months of personal use on real DevOps workflows, agent integrations, and deployment pipelines.

Use Cases

DevOps and Infrastructure Teams: Before piping kubectl, docker, or helm output to Claude for debugging and suggestions, run it through Lowfat. Reduce token costs by 80-95%. Same advice, 1/10th the price.

Autonomous Agents: Agents that make decisions from system logs, error traces, deployment output, or cluster state can now run 10x more queries for the same API budget. Or run the same queries at 1/10th the cost.

CI/CD Pipelines: Agent-driven CI/CD (tests, deployments, rollbacks) generate massive logs. Lowfat filters them before feeding to the LLM. Faster decisions, lower token spend.

Log Analysis and Monitoring: Feeding raw application logs to Claude or ChatGPT for analysis? Lowfat extracts the signal, removes duplicates and timestamps, keeps the error messages and stack traces that matter.

Git and Code Analysis: Large git logs, diff output, or code analysis results become manageable. Lowfat strips the boilerplate, keeps the semantics your LLM needs.

Getting Started

Installation: Download the binary from GitHub releases or build from source.

$ brew install zdk/tap/lowfat
# or
$ go install github.com/zdk/lowfat@latest
      

As a shell wrapper:

$ lowfat kubectl get pods | xargs -I {} curl http://api.example.com -d {}

As an agent hook (Anthropic SDK):

$ export ANTHROPIC_PRE_API_HOOK=lowfat
# Lowfat automatically filters all CLI output before sending to Claude
      

Custom plugin for your tool:

$ cat ~/.lowfat/filters/mytool.yaml
command: my-verbose-tool
strip_fields:
  - timestamp
  - debug_*
  - metadata.*
keep_patterns:
  - error
  - warning
  - result
      

Then run lowfat my-verbose-tool args and it automatically applies your filters.

Pricing

Lowfat is free and open source under the MIT license. Use it anywhere, customize it, redistribute it. No commercial license required.

Commercial Support: For teams needing dedicated support, custom filters, or SLA commitments, contact us on GitHub for commercial licensing options.

Why Lowfat Wins

Economics: 91.8% token cost reduction proven real. ROI is immediate for any team making frequent LLM API calls.
Simplicity: Single binary. No config required for common tools. Zero dependencies. Works offline.
Extensibility: Plugin system lets you customize per your workflows. But defaults are smart for 80% of DevOps use cases.
Developer-Native: Open source. Local-first. No vendor lock-in. Built by developers, for developers.
Real Tracking: Cost savings are tracked and visible. You know exactly what you're saving. Not theoretical, not estimated, actual.