Pack
Deterministically pack context before it goes on the wire — stop re-sending the same boilerplate, file contents, and system prompts.
TokenPak — local LLM proxy
Local. Context-lean. Measurable.
TokenPak is a local proxy that packs your LLM context before it hits the API — fewer repeated tokens on the wire, more reuse opportunities, and savings you can measure per-request.
TokenPak doesn't replace your AI stack.
It adds the missing logistics layer: packing, routing, reusable context, guardrails, orchestration, and per-request records.
Six capabilities
Route
Send each request to the right model, with fallback rules you control.
Reuse
Recall and reuse context from local PAKs instead of rebuilding it every session.
Guard
Spend and safety guardrails run as a side-channel gatehouse across every request.
Dispatch
Coordinate scoped, multi-step and multi-agent work. Preview (alpha) — not yet in a published release.
Record
Every request leaves a local record — what was sent, reused, observed, and charged.
How a request flows
One linear path, packed and recorded locally before it leaves your machine.
- Raw context
- Packing Station
- Route
- Send
- Provider
- Record
Reuse feed: PAKs feed into the Packing Station — recall and reuse context instead of rebuilding it.
Gatehouse (Guard): spend and safety guardrails run as a side-channel overlay before Send.
Where TokenPak fits
TokenPak overlaps with tools you may already run — and adds a layer they don't. Run them together.
How it sits alongside gateways, observability tools, and MCP
-
Your agent
Claude Code
Cursor · Cline
-
TokenPak (local)
pack · measure
guard · record
-
Your gateway
LiteLLM /
OpenRouter / …
-
Model provider
Anthropic · OpenAI
Google Gemini
Dispatch Center dashboard (preview) — orchestration for multi-step and multi-agent work, around the flow.
| Tool type | Overlaps on | TokenPak adds | Together? |
|---|---|---|---|
| Gateways / routers | routing | deterministic context packing + reusable PAKs | Run both. |
| Observability tools | measurement | pre-send packing + per-request context-level attribution | Run both. |
| MCP-based workflows | ecosystem coordination | a semantic contract (TIP) for packing, routing, cost, telemetry | Composes. |
Savings
TokenPak avoids re-sending context your tools already sent. We don't publish savings numbers yet — receipt-backed measurements will land with the benchmark work. Until then, every request leaves a local record you can check yourself.
Open source & Pro
TokenPak's core is open source (Apache-2.0). Pro adds team dashboards, advanced routing, and enterprise controls.
Pro is delivered as the tokenpak-paid package via a separate index.
Latest release
latest
### Added - **Codex receipt-only launch mode.** `--receipt-only` (requiring `--receipt-out` and `--run-id`, mutually exclusive with `--budget` and `--install-only`) lets a launch emit accounting receipts without installing the TokenPak mechanism, and launches without a request body now produce **no-body accounting receipts** so accounting stays truthful instead of silently dropping those events. - **Canonical staging-to-public promotion tooling.** `scripts/promote-staging-to-public.sh` and `scripts/promotion-drift-report.sh` codify the promotion train between the staging and pu
Documentation
starter
Getting Started
Install TokenPak and see savings in one command.
starter
Installation
pip install options, OS notes, troubleshooting install failures.
starter
Claude Code guide
Route Claude Code through TokenPak with one environment variable.
More at docs.tokenpak.ai.
Try it locally.
pip install tokenpak && tokenpak setup — then point your AI client at the local proxy. No cloud component; credentials stay in your environment and provider flow.