TokenShift

Most engineering orgs have no view into what their coding agents are doing: no sense of where the token spend goes, no idea what the agent actually touches on a developer’s machine.

TokenShift closes that gap from the endpoint up. It hooks into the coding agent harnesses your team already runs — Claude Code, Cursor, GitHub Copilot CLI, and Codex — and does two things, every time an agent runs a command:

• Cuts the token spend. A local compression engine trims command output before it ever reaches the model — shorter git diffs, test runs, kubectl dumps — without changing what the agent can see or do. Nothing is silently dropped: the full output stays on the machine, and the agent can ask for it back by name.
• Makes the fleet legible. Every tool call becomes a signal — what ran, on which harness, how many tokens it cost, how much was saved — redacted on the device, encrypted before it ships, visible on the dashboard.

Install is one command. The binary itself runs per call, invoked directly by the agent’s own hook system — nothing persistent to babysit. The one background helper it spins up (a small on-device model that classifies what a prompt is about) lives only as long as you’re actively using the harness and shuts itself down after an hour of no activity. tokenshift install detects which harnesses are on the machine and wires the rest in one step.

That’s the whole shape of it: see it, shrink it — one lightweight binary dropped into a workflow your engineers already use. No retraining, no new muscle memory.

Where to next

Install in 60 seconds

Drop TokenShift into Claude Code, Cursor, GitHub Copilot CLI, or Codex on a single machine and see it work. Quickstart →

What it does

The architecture in one page — proxy, hook, recovery, telemetry. How it works →

For your security team

What goes on the wire, how it’s encrypted, and what stays on the device. Security overview →

Every field, every tier

The full data contract — what’s collected, what’s not, and how long it’s kept. Data contract →

Scored, harness by harness

How completely each agent is covered today — one score per harness, per OS. Harness coverage →