tokenpak 1.17.0

Cut LLM costs with deterministic context compression, smart routing, and cost tracking

TokenPak — local context packing with measurable receipts

The open logistics layer for AI context.

TokenPak is a local proxy that packs AI requests before they ship and records what changed. It helps developers reduce repeated context on eligible routes, without moving prompts or credentials into a TokenPak cloud service.

---

30-second demo

python -m pip install tokenpak
tokenpak demo

┌──────────────────────────────────────────────────────┐
│  TokenPak — Offline Fixture Demo                     │
├──────────────────────────────────────────────────────┤
│  Scenario              DevOps agent (config + logs)  │
│  Data source                built-in sample fixture  │
│  Savings drivers                      dedup + alias  │
├──────────────────────────────────────────────────────┤
│  Original                                747 tokens  │
│  Compressed                              502 tokens  │
│  Fixture delta                  245 tokens  (32.8%)  │
│  Fixture cost delta            $0.00073 per fixture  │
│  Receipt status               not a savings receipt  │
├──────────────────────────────────────────────────────┤
│  Stages: dedup, alias, segmentize, directives        │
└──────────────────────────────────────────────────────┘

Then verify:

tokenpak --version

tokenpak 1.17.0

The offline fixture is illustrative, not a measured savings receipt. Token counts vary by route and workload. For the signposted measured path, run a supported authenticated client through the proxy and follow the first-receipt guide.

---

Works with

Claude Code · Cursor · Cline · Continue.dev · Aider · OpenAI SDK · Anthropic SDK · LiteLLM · Codex

Run tokenpak integrate to see supported-client setup guidance.

---

Install

python -m pip install tokenpak

Requires Python 3.10+. See the install guide for isolated-install options and quickstart for client setup.

---

What's included

• Context packing — reduces repeated eligible context before provider send; inspect measured results with tokenpak savings.
• Local proxy and client integration — route supported clients through one local service without changing application code.
• Spend Guard — a pre-send circuit breaker that can block configured runaway requests before the provider call.
• Local cost and receipt records — inspect measured activity by model, session, and route.
• Vault indexing and semantic search — index a codebase and search it locally.
• Compression recipes — 50 configurable YAML recipes for supported flows.

---

How it works

Your AI client sends provider-compatible requests to the TokenPak proxy on 127.0.0.1. TokenPak compresses eligible context and evaluates configured Spend Guard limits before forwarding to the selected upstream provider, then records the request result locally.

AI client -> TokenPak proxy -> upstream provider
                   |
                   +-> local receipt and cost records

---

Documentation

• Quickstart
• API reference
• Architecture
• First measured receipt
• Spend Guard
• Troubleshooting
• Runnable examples

---

Runnable examples

The repository examples are not bundled inside the PyPI wheel. Read the examples/README.md, or clone the source and run the local, credential-free compression example:

git clone https://github.com/tokenpak/tokenpak.git
cd tokenpak
python -m pip install -U tokenpak
python examples/basic_compression.py

---

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.

---

License

The TokenPak open-source core is Apache-2.0 licensed. TokenPak Pro and hosted services are proprietary. See LICENSE and SECURITY.md.