Relay · control plane (v3 mock)

Everything's on the wing

Four accounts enrolled, the proxy's up, and nothing live is near its limit. Vega needs a fresh sign-in; the other three are fine.

Accounts enrolled

2 healthy · 1 near cap · 1 needs sign-in

Tokens compressed today

lossless · ~34% of input

Spend this month

≥ partial; 2 records unpriced

Account near cap

Nimbus · 91% of 5-hour

Accounts

Connections

Claude Code

ANTHROPIC_BASE_URL → :8787

Connected

Codex

Responses → :8787

The proxy port is sticky but it self-heals. If it moves, Relay rewrites each agent to the port it actually bound.

Live now

Flow

How each bound project's requests travel through compression to the account it's routing to right now. Faint lines are the map; a line lights up and carries packets only while a folder is actively sending.

Healthy Near reserve Capped

34%

Compression

lossless · 1.24M saved

Live. Accounts, weekly usage and compression totals are read straight from this Relay. A project moves to the next account in its chain once the one it's on crosses its reserve.

Accounts

Every enrolled login, and how much of its window is left. Relay moves new work down a project's chain before any of these hits its reserve, so a session never walks into a wall.

Enrolled · live usage

polling every 60s

Credits & gateways

OpenRouter

gateway · pay-as-you-go

$12.40

Vercel AI Gateway

gateway · key required

$8.10

Cursor

API credit + Auto

$40.00

Auto 72%

Bindings

A binding points a project folder at an ordered chain of accounts and pools. Work starts at the top and only drops to the next when an account crosses its reserve. Conversations already running stay on their warm account, so you keep the prompt cache.

Default

Any folder without an explicit binding

all accounts, by weekly headroom

Account pools

Personal

Atlas · Nimbus; reclaims earlier members when they recover

Project bindings

Spend

Estimated cost of routed and gateway traffic this month, split by model and project. Subscription usage is metered but not billed here; these figures are the paid slices only.

This month

Last 30 days

All time

This month

≥ 2 records unpriced

OpenRouter

gateway spend

Vercel AI Gateway

gateway spend

Compression avoided an estimated $3.90 of gateway spend this month by never sending trimmed context.

By model · this month

By project · this month

Activity

A restart-safe, content-free record of what Relay did: account switches, limits reached, compression, and proxy recovery. No prompts or responses are ever stored here.

Recent events

Today

Atlas → Nimbus, switched ahead of the 5-hour reserve (predictive · re-warms ~4k)

09:38

Compressed 312K tokens of tool output, lossless

09:31

Nimbus reached its 5-hour limit; new work held for eligible accounts

08:57

Yesterday

Proxy recovered; re-bound :8787 and rewrote connected agents

18:12

Live sessions

~/Dev/perch on Atlas · opus-4.8 working

saved 18.6K tokens this session

Intelligence

Two optional layers, both off by default and fully auditable. Routing sends each task to the model that suits it; compression trims what gets sent without changing meaning. The essentials are here; open a section for the full rulebook.

Task routing

Relay classifies each prompt and sends it to the model in your rulebook. Off means every request runs on its own model, unchanged.

Router source

The small, cheap model that reads each prompt to classify it.

Active profile

A preset that sets the whole rulebook at once. Cache-first keeps conversations on their warm account.

Cache-first

Rules for

Apply globally, or override for a single project.

Ask before routing

Confirm each switch before it happens, rather than routing silently.

Task → model

Task typeRoutes toAlways

Design / UIAsk

RefactorAsk

Bulk editOn

DebugOn

DocsAsk

PlanningAsk

TestsAsk

Data / JSONAsk

Model roles

Route by role

Plan mode

While plan mode is active

Implement

When you leave plan mode

Subagent

A subagent's first turn

Review

Unavailable — no automatic signal

—

Cache-first Active

Strong evidence · keep conversations on their warm account

Applied

Economy

Provisional · cheapest capable model per task

Balanced

Provisional · quality on hard tasks, thrift elsewhere

Quality

Provisional · best available model everywhere

Planner → Builder Experiment

Staged: plan on opus, build on sonnet

High-value deliberation Experiment

Panel + judge on the hardest calls

Lower-cost diversity Experiment

Spread work across cheaper providers

Experiments have measured lift not yet established. Each shows its cache-rewarm, return-trip and multi-model-spend caveats before it applies, and any preset change can be undone.

Ask several models the same question, then a judge synthesises one answer. Read-only: ordinary model calls, no tools or web, 120-second timeout.

Conclusion

Research answer

Implementation plan

Code review

Decision compare

Question, decision, or plan to deliberate on…

Panel 1

Panel 2

Judge

Output limit

Warn if >

est. $0.18

Remembered "always" rules

Debug → always claude · opus-4.8

Bulk edit → always claude · haiku-4.5

Recent decisions

TimePromptModelHow

09:38"tidy the flow legend spacing"sonnet-5Auto

09:12"why does the proxy port move"opus-4.8Confirmed

08:44account switch · re-warms ~4kAtlas→NimbusHanded off

Mid-task hand-off Advanced

GLM-4.6 "drifted into design work" claude · sonnet-5 next turn

Confirmation style

In-CLI shows a numbered prompt in your terminal ([1] yes [2] keep [3] pick… [a] always); Seamless uses a native macOS notification you can act on without leaving the editor.

Best execution

Treat your chosen route as a quality grade, then let Relay find the cheapest way to keep it. Advisory for now: it explains what it would choose and why, and never acts on a warm conversation.

Advisory

Reference grade

The quality bar every cheaper candidate must clear. Price can never buy its way past this.

Quality floor

Of the tasks the reference solves, the candidate's conservative lower bound must retain at least this share.

Latest decision · explained from the record

Stayed on sol · high. The cheaper candidate met the benchmark floor but didn't clear the projected 92k-token cache re-warm plus uncertainty margin. Evidence: 148 matched local episodes + benchmark prior · policy best-exec-17 · updated 3d ago.

Counterfactual: would switch after ~3 more warm turns at current prices. Reconsidered at the next compaction or plan-to-build boundary.

Was this right?

Feedback joins the decision by an opaque id — a bounded metric, never your prompt. It tunes which candidates clear the floor for your workload, the same loop Ship's feedback API closes for theirs.

Compression

Trim historical tool output before it's sent. Lossless is byte-reversible; higher modes normalise or drop low-signal lines.

0tokens saved this session

Roughly 34% of what you'd have sent · 41 blocks · estimated (Anthropic's tokenizer isn't public).

Command logs

JSON payloads

Search output

Diffs

Protected, safe by default

Read / Edit / Write & file reads

Native, shell and MCP reads stay exact; edits need the original bytes.

Error output

Tracebacks kept verbatim so the model can recover.

System & tool definitions

Always exact.

Cached prefix

The already-cached prompt prefix is never rewritten.

Prompt-cache impact · 24h

86%

cache hits

≈4%

new-account miss

2.1M

Claude input

Cache hit is cache-read input over all Claude input. New-account miss attributes the prior-context portion re-written on the first completed turn after a switch; ≈ marks the unobservable same-account counterfactual.

1.8M

treatment cache reads

312

warm hits

profile changes held

Control vs treatment cohortlast 7 days

Control · 5% Off holdout

142 turns · avg 4.1s · repeated reads 18

Treatment · compression on

2.6K turns · avg 3.9s · repeated reads 21

Estimated $3.90 gateway spend avoided · API-equivalent priced-turn cost held flat.

Pinned Off holdout

Keep a slice of turns uncompressed to measure real lift.

Both layers fail open. If anything's uncertain, Relay sends the original, uncompressed, on the client's own model. Nothing here can make a request fail.

Settings

Connections, the proxy, polling, alerts, and appearance. The same controls back the ⌘, preferences window.

Agents & proxy

Claude Code

ANTHROPIC_BASE_URL → 127.0.0.1:8787

Connected

Codex

Responses → 127.0.0.1:8787 · ~/.codex/config.toml

Providers & gateways

OpenRouter

318 models · updated 6m ago

Connected

Vercel AI Gateway

API key

Cursor

API credit + Auto · login

Connected

Proxy

Port

8787

Auto-pick a free port

Falls forward if the port is taken.

Always on

Polling

Usage interval

60s

Balance interval

60s

Usage alerts

Notify when 5-hour usage crosses

85%

Notify when weekly usage crosses

90%

Alert on low gateway balance

Below this remaining credit.

General

Appearance

Launch Relay at login

Menu-bar cycle interval

How fast the tray meter rotates between active accounts.

About

Relay

3.0.0 (mock)

Proxy

127.0.0.1:8787running

Notifications

authorised

Overview

Everything's on the wing

Flow

Accounts

Bindings

Spend

Activity

Intelligence

Settings