Package
tokenpak
The Python package — proxy, companion, CLI, integrations. Published to PyPI; Apache 2.0. What a user installs.
About
Up to 90%+ savings on direct API/CLI and other favorable uncached workloads, with zero code changes.
Observed savings vary by integration path and provider-side cache behavior. Provider-cached flows (Claude Code and similar) may show lower incremental gains because TokenPak only affects the user-controlled portion of the token pool.
TokenPak is the local layer between a developer's LLM client and the provider. It compresses context before the wire, routes requests intelligently, and tells the user what was saved — without touching their credentials or their code.
Product identity
TokenPak is a local proxy that compresses context before it hits the LLM API. That sentence goes in the README, the homepage, and every pitch. The identity is deliberate: local-first (no cloud to fail, no credentials to store), proxy-shaped (drops in under the tools you already use), compression-focused (one job, done deterministically).
Compression is one job, not the whole story. Routing, caching, telemetry, and a local dashboard all ship in the same package — because a team running agents in production needs all four at the same layer.
TIP-1.0 — the TokenPak Integration Protocol
TokenPak is the reference implementation of the TokenPak Integration Protocol, version 1.0 — a small specification that defines what TokenPak-compatible behavior means: canonical headers, metadata fields, telemetry event schemas, capability labels, manifest shapes. TIP is a profile/spec, not a transport; the normative text lives in the docs repo and the machine-readable schemas live in the registry repo.
TokenPak adopts MCP (Model Context Protocol) as its default control-plane substrate. MCP and TIP are different layers: MCP is the transport; TIP is the semantics.
Where things live
Four public surfaces. Each owns one concern.
Docs
docs.tokenpak.ai
MkDocs-rendered user docs + TIP-1.0 normative spec. Sourced from the tokenpak/docs repo.
Registry
tokenpak/registry
Machine-readable TIP-1.0 schemas, manifest schemas, capability catalog, adapter + plugin manifests. Published as tokenpak-tip-validator on PyPI.
Site
tokenpak.ai
This page — the public marketing surface. Astro static site; source in tokenpak/site.
Source on github.com/tokenpak · Published on PyPI · Docs at docs.tokenpak.ai.
Ready to try it?
pip install tokenpak. Works with the LLM clients you already use.