For people who live in Claude Code & Codex

Every model you pay for, working as one.

Relay runs your Claude and Codex subscriptions, and your own API keys, as one endpoint on your Mac. It switches accounts before limits hit, routes each task to the model that's good enough, and squeezes every token on the way through.

Runs on your machine Keys stay in the Keychain Works with subscriptions you already pay for

The wall

You're mid-task. Claude isn't.

You're deep in a task, the agent's mid-thought, and Claude tells you you've hit your five-hour limit. Come back in three hours.

You've probably got another account sitting right there with plenty left. You just can't reach it without logging out, dropping your session, and starting the whole thing again. Relay reaches it for you, between turns, so nothing breaks mid-request.

claude · ~/projects/api
implementing auth middleware…
·34 files read · 12k tokens out
5-hour limit reached. Try again at 6:12 pm.
relay → nimbus has 82% left · switching between turns
continuing on nimbus · same session, nothing lost
middleware done, writing tests…
atlasreserve hit
5-hour window · resets 6:12 pm · reserve 85%
nimbusactive
5-hour window · resets 7:40 pm
orbitnext
weekly window · 96% left

Multi-account

It switches before you run out, not after.

Relay watches every account's five-hour and weekly windows live. When the one you're on gets close, it quietly moves the next turn to another account. No limit, no error, no pause; the session just carries on.

Leave some in the tank. Sharing an account with your team? Set a reserve (say 85%) and Relay stops there and moves on.

When everything's spent, you still don't stop. Relay keeps the session going on another model: GLM or Gemini through OpenRouter, or Codex on GPT. It's not Claude, and Relay shows you exactly which model you've dropped to. But it beats a hard stop and a three-hour wait.

A real Mac app, not a browser tab. Menubar meter, statusline, per-project bindings. Quiet until it matters.

Four ways Relay stretches every token

Between your prompt and the model, Relay earns its keep.

Every request passes through four disciplines. Each one either cuts the cost or raises the quality, and each one explains itself.

01 · Compression

Long sessions, short bills.

An agent re-reads its whole conversation on every turn. Relay's compression layer trims what the model no longer needs, without ever breaking the provider's prompt cache, so the savings are real, not shuffled around.

128k input tokens per turn · watch them shrink
(illustrative session)
02 · Auto-routing

Not every task deserves the priciest model.

Profiles are labelled with the evidence behind them: benchmark results, sample sizes, freshness. Relay stays on the warm route unless a switch clearly wins after cache costs. Preview before apply, one-step undo.

good enough is the target, never “cheapest at any cost”
03 · Fusion

Two opinions. One judge. Your call.

For plans and reviews that are expensive to get wrong. You see who'll be asked and what it costs before anything is sent, and nothing gets built until you approve the result.

N + 1 calls, disclosed up front · opt-in, never silent
04 · Guard

The wall never reaches your session.

Reserves, pre-emptive switching, and honest model fallback when every account is spent. Every hand-off happens at a turn boundary, so nothing breaks mid-request.

0 sessions lost to a usage window

Everything in the box

A control plane, not a dashboard.

Live usage meters

Five-hour and weekly windows per account, in your menubar and statusline. Relay pings you before one gets tight.

menubar · statusline · alerts

Routing profiles

Economy, Balanced, Quality; each labelled with the evidence behind it. Preview before apply, one-step undo, and manual edits always win.

evidence-labelled · cache-aware

Fusion deliberation

Ask two models independently, let a judge compare, approve before anything gets built. Costs shown before a single call is made.

opt-in · read-only until approved

Compression layer

Trims what the model no longer needs from long sessions; cache-safe by design, so a smaller prompt never costs you a cache re-read.

cache-safe · always reversible

Spend ledger

What you actually used, per model and per project, with cached and fresh tokens counted separately, so fallback models never surprise you.

per model · per project · cache lanes

Skills that trim output

Reusable instructions that keep answers short where short is right, because output tokens are the ones that quietly eat a budget.

output discipline · per-project

Local first

Everything happens on your machine.

Relay is a loopback proxy: your editor talks to 127.0.0.1, and Relay talks only to the providers you've connected. Your prompts, your code, and your keys never route through anyone else's cloud, including ours.

keys

OAuth tokens and API keys live in the macOS Keychain. They never appear in logs, state files, or exports.

prompts

Never logged, never stored, never used to train anything. Relay's local scorebook keeps token counts and outcomes, not content.

network

Binds to localhost only. Block every Relay host at your firewall and the app keeps working.

sharing

Off by default. If you opt in, only thresholded, content-free totals leave, and you can preview exactly what they contain.

claude code
relay :8787
codex
relay :8787
relay
your providers only
your mac · keys in keychain, prompts stay home

One idea, three shapes

On your Mac today. On your server and across your team next.

Available now

Relay App

The native macOS app. Multi-account switching, routing, fusion, compression, spend tracking; all local, all yours.

Your subscriptions & API keys
Menubar, statusline, CLI
Works with all cloud services blocked
Coming

Relay Cloud

The same endpoint, hosted: for CI, remote boxes, and machines that aren't your Mac. Bring your own keys; they're envelope-encrypted with per-tenant locks.

Same APIs: Messages, Responses, Chat
Your keys, encrypted at rest
Requests processed, never retained
Preview the cloud app →
Coming

Relay Teams

Pool the subscriptions your team already pays for. Reserves keep everyone working; shared routing evidence makes every seat smarter.

Pooled accounts with per-person reserves
Per-person caps, team spend by project
Shared evidence, never shared prompts
See team management →

Subscription tokens are the cheapest tokens.

A month of heavy agent use, three ways; then Relay stacks the rest.

Metered API pricingpay per token, list price
$100
Pooled subscriptionsthe plans you already pay for, via Relay
$38
+ routing, compression & skillscheaper routes, smaller prompts, shorter answers
$24

illustrative comparison · your mix of models, tasks and caching will move these numbers

Optional companion · built on Ghostty

Tandem
the terminal that travels

One live terminal. Every machine you own, and the people you trust.

Tandem is a native macOS terminal where two or more Macs share the same session: same cells, same cursor, same scrollback, one authoritative shell. Your agent sessions stop being trapped in the laptop that started them.

runners

Start a session on a Relay cloud runner; it keeps working after your lid closes.

hand-off

Move a Git-backed session between Macs (agent conversation and background jobs included) and pick up exactly where you left off.

mirroring

A session on one Mac appears as an ordinary, quietly cloud-marked tab on your other Mac.

sharing

Share a live terminal read-only, or grant short write turns; one writer at a time, revocable instantly.

macOS · native, not Electron iPhone companion Web viewer tandem.fledgeling.app ↗
api-refactor · claude deploy-scripts motif · main
# started on robin's MacBook · now hosted on runner east-1 ▸ claude · refactoring the auth middleware… ✓ 12 tests passing · branch api-refactor ahead 3 ⧉ mira joined read-only · requested a write turn grant 5m · one writer at a time, revoke any time
host east-1writer mira · 4:12 viewers 2epoch 17

Pricing

Buy the app. The rest is optional.

App
$59 one-time

The full native app, yours forever. Every feature on this page that says “available now.”

  • Unlimited accounts & projects
  • Routing, fusion, compression
  • Free updates through v2
Download for Mac
Teams
$18 /seat/mo

Pooled subscriptions beat metered API pricing; Relay makes them work like one well-behaved endpoint.

  • Pooled accounts, per-person reserves
  • Team spend & routing evidence
  • Admin controls & audit
Join the waitlist
Cloud
$12 /mo

The hosted endpoint for CI and remote machines. Your keys, encrypted; your requests, never retained.

  • Messages, Responses & Chat APIs
  • Encrypted BYOK vault
  • Usage metering & caps
Join the waitlist

beta pricing, illustrative · cloud & teams ship after the app

Relay

Stop planning your day around a usage window.

Download for Mac