Pruner

A zero-config local proxy that cuts your Claude Code API bill by 20–70%.

Pruner sits between Claude Code and the Anthropic API, automatically optimizing every request in real time. It trims redundant context, injects prompt caching, and reports exactly how much you saved — verified by Anthropic's own tokenizer.

────────────────────────────────────────────────────────
 Pruner  #4  31,204→18,940 tok✓  -39.3%  $0.037 │ ⚡ 48,570 cached✓  $0.131 │ Σ $0.412
────────────────────────────────────────────────────────

---

Table of Contents

• How it works
• Security
• Install
• Usage
• Configuration
• Architecture
• Contributing
• License

---

How it works

When you run pruner, it:

1. Starts a local HTTP proxy on 127.0.0.1:7777
2. Sets ANTHROPIC_BASE_URL=http://127.0.0.1:7777 in the environment
3. Spawns the real claude CLI with all your original flags

Every API request passes through three optimization stages before reaching Anthropic:

Stage	What it does
Context pruning	Removes old messages beyond maxMessages. Inserts a placeholder so Claude understands the gap. Truncates oversized tool outputs.
Prompt cache injection	Automatically adds cache_control: { type: "ephemeral" } to system prompts over 1,024 tokens. Cache reads cost 10× less than regular input ($0.30 vs $3.00 per million tokens).
Verified savings	Calls /v1/messages/count_tokens in parallel (zero latency) to get Anthropic's own token count. Reports actual usage.input_tokens from each response.

Nothing is sent to any server other than api.anthropic.com. Verify it yourself:

pruner --debug
# [debug] → api.anthropic.com:443  POST /v1/messages
# [debug] ✗ no other outbound connections

sudo lsof -i -n -P | grep pruner  # independent verification

---

Security

Pruner is a local-only proxy. Your code, prompts, and API key never leave your machine except to go directly to Anthropic — the same destination they'd reach without Pruner.

What Pruner does and does not do

	Detail
✅ Binds only to localhost	The proxy listens on 127.0.0.1:7777 — not reachable from your network or the internet
✅ Only connects to api.anthropic.com	Zero telemetry, no analytics, no third-party servers
✅ API key is never stored	Forwarded in-memory, transparently, to Anthropic — identical to what Claude CLI does
✅ No data collection	Pruner has no backend, no accounts, no servers that receive your data
✅ Open source (MIT)	Every line of code is on GitHub — read it, audit it, compile it yourself
✅ Signed binaries	The install script verifies a SHA-256 checksum before installing

Verify it yourself

You don't have to take our word for it. Use --debug mode to see every outbound connection Pruner makes:

pruner --debug
# [debug] → api.anthropic.com:443  POST /v1/messages
# [debug] ✗ no other outbound connections

Or use your OS independently:

# macOS / Linux — see every socket opened by the pruner process
sudo lsof -i -n -P | grep pruner

You should see exactly one remote address: api.anthropic.com:443.

Build from source

If you don't want to trust a pre-built binary, building from source takes under a minute:

git clone https://github.com/OneGoToAI/Pruner.git
cd Pruner
npm install
npm run build:bin:mac-arm   # or mac-x64 / linux

The compiled binary is bit-for-bit identical to the one in the release — same build command, same Bun version (pinned in CI).

---

Install

macOS / Linux — one line

curl -fsSL https://raw.githubusercontent.com/OneGoToAI/Pruner/main/install.sh | bash

The script detects your OS and architecture, downloads the correct binary, verifies its SHA-256 checksum, and installs to /usr/local/bin.

macOS — Homebrew

brew install OneGoToAI/tap/pruner

Build from source

Requires Bun ≥ 1.0 and Node.js ≥ 18.

git clone https://github.com/OneGoToAI/Pruner.git
cd Pruner
npm install
npm run build          # TypeScript → dist/
node dist/index.js     # run directly

# Or compile a self-contained binary
npm run build:bin:mac-arm   # Apple Silicon
npm run build:bin:mac-x64   # Intel Mac
npm run build:bin:linux      # Linux x64

---

Usage

Replace claude with pruner. Every Claude CLI flag is passed through unchanged.

# Start an interactive session
pruner

# Resume the last conversation
pruner --resume

# Run a one-shot prompt
pruner -p "refactor the auth module for testability"

# Verify Pruner's network activity
pruner --debug

When you exit, a session report is printed:

────────────────────────────────────────────────────────
 Pruner  Session Report  ✓ verified
────────────────────────────────────────────────────────
 Requests                        12
 Original tokens            128,432
 After pruning               31,204
 Pruning saved           75.7%  $0.29
 Cache hit tokens         48,570 ✓
 Cache saved                   $0.13
────────────────────────────────────────────────────────
 Total saved                   $0.42
 Duration                       183s
────────────────────────────────────────────────────────

Numbers marked ✓ verified come directly from Anthropic's API. Numbers marked ~estimated use tiktoken as a fallback (within ~10%).

Commands

Command	Description
pruner	Start Claude with cost optimization
pruner --debug	Show every outbound connection
pruner config	Open ~/.pruner/config.json in your editor
pruner reset	Reset session statistics
pruner <any claude flag>	All Claude CLI flags pass through

---

Configuration

Run pruner config to open ~/.pruner/config.json. Changes take effect immediately — no restart needed.

{
  "proxyPort": 7777,
  "optimizer": {
    "enablePromptCache": true,
    "enableContextPruning": true,
    "enableTruncation": true,
    "accurateTokenCounting": true,
    "maxMessages": 20,
    "maxToolOutputChars": 3000
  },
  "pricing": {
    "inputPerMillion": 3.0,
    "outputPerMillion": 15.0,
    "cacheWritePerMillion": 3.75,
    "cacheReadPerMillion": 0.3
  }
}

Option	Default	Description
proxyPort	7777	Local proxy port. Increments automatically if occupied.
enablePromptCache	true	Inject cache_control on large system prompts.
enableContextPruning	true	Trim old messages past maxMessages.
enableTruncation	true	Cap individual tool outputs at maxToolOutputChars.
accurateTokenCounting	true	Use Anthropic's count_tokens API for exact figures.
maxMessages	20	Maximum messages to keep. First message is always preserved.
maxToolOutputChars	3000	Characters per tool result before truncation.
pricing.*	claude-3-5-sonnet rates	Override if you use different models or have custom pricing.

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│  pruner (src/index.ts)                                       │
│  ├── Checks for 'claude' CLI                                 │
│  ├── Starts local proxy on 127.0.0.1:7777                   │
│  └── Spawns claude with ANTHROPIC_BASE_URL injected          │
└─────────────────────────┬───────────────────────────────────┘
                          │ all API traffic
┌─────────────────────────▼───────────────────────────────────┐
│  AnthropicProxy (src/proxy.ts)                               │
│                                                              │
│  Incoming request                                            │
│  ├── [1] optimize()         src/optimizer/index.ts           │
│  │       ├── injectPromptCache()   optimizer/cache.ts        │
│  │       └── pruneContext()        optimizer/pruner.ts       │
│  │             └── truncateLargeContent()  truncate.ts       │
│  ├── [2] fetchExactTokenCount()   stats/counter.ts  ─┐      │
│  └── [3] undiciRequest() → api.anthropic.com          │ parallel
│                                                        │      │
│  Response                                              │      │
│  ├── parse usage.input_tokens (verified compTokens)   │      │
│  ├── await count_tokens result  ◄─────────────────────┘      │
│  ├── recordRequest()        stats/session.ts                 │
│  └── printRequestLog()                                       │
└─────────────────────────────────────────────────────────────┘

Key files:

File	Purpose
src/index.ts	CLI entry point, process management
src/proxy.ts	Fastify proxy server, request/response pipeline
src/config.ts	~/.pruner/config.json with hot-reload
src/optimizer/cache.ts	Prompt cache injection
src/optimizer/pruner.ts	Context pruning and message cap
src/optimizer/truncate.ts	Large content head+tail truncation
src/stats/counter.ts	tiktoken + Anthropic count_tokens API
src/stats/session.ts	In-memory session statistics

---

Contributing

Contributions are welcome. Please read CONTRIBUTING.md first.

# Setup
git clone https://github.com/OneGoToAI/Pruner.git
cd Pruner
npm install

# Development (runs TypeScript directly via tsx)
npm run dev

# Tests (32 unit tests, Vitest)
npm test

# Lint
npm run lint

# Type check
npm run build

Opening a PR? Please:

• Add or update tests for any behavior change
• Run npm test and npm run build before pushing
• Keep commits focused; one logical change per PR

For larger changes (new features, architectural changes), please open an issue first to discuss the approach.

---

License

MIT — see LICENSE.

---

Acknowledgements

• Fastify — the proxy server
• undici — HTTP client
• tiktoken — fallback token counting
• Bun — binary compilation
• Anthropic — for the count_tokens API that makes verified savings possible