Co-ordinate Get the launch emailLaunch email

The demo city in the Retrowave design: an island of lit towers at dusk on a neon grid sea, under a striped sun. Each building is a project with its name on a sign; lit windows are chats.

Every Claude and Codex agent you run, in one city.

See what they do, and what they waste.

Co-ordinate is a local dashboard for Claude Code and Codex. Each project is a building and each chat a lit window, so you see who is working, who is waiting for you, and which chats are quietly costing you. Watching it costs zero tokens.

Get the launch email See all six designs

Free and open source (MIT) for one person. One email when it is out.

See it as

Demo city: 16 made-up projects, drawn by the real Co-ordinate.

The same city in six designs.

Six designs in twenty seconds: Retrowave, Fantasy, Wasteland, Edo, Stone Age and Lunar.

The same 16 projects and the same signs in each; pick a design in the app's Design row, or above.

Read every agent in one look.

You don't scan forty rows. You look at a skyline, and the trouble stands out.

Close-up of the demo city at night: a tower for the Mobile App project with lit windows, its sign above it, a red window and a red roof beacon on the building beside it.
  1. A building is a project. Its shape comes from the project; its colour stays the same in every design.
  2. The sign names the project and says how it is doing: running, waiting for you, blocked.
  3. Each lit window is a chat. A teal sill marks a Codex chat (Cursor and Gemini CLI get their own teal trim); the rest are Claude.
  4. Red means blocked. That agent is stuck until someone helps.
  5. The roof beacon calls you. A blocked chat or a warning lights it.
  • Busy
  • Idle
  • Waiting for you
  • Blocked
  • Asleep
  • Codex chat

Each building is a project with its HQ: a small local page that reads the project's chats and its STATUS.md. A line at the top of the city says what waits for you, and where. Click a window to open its chat, a building to open its project; Board, List and HQ show the same city as cards and tables.

Act only where you're needed.

Approvals is one list of what waits for your answer: chats, decisions, permission prompts, paused runs and new HQs. Notices, such as tips and cap warnings, come after it.

  • Waiting for you is one count, in the city's top line, on each project's sign and in Approvals. Answer a waiting chat from its row.
  • Roles and tasks for Claude and Codex agents live in Command. A task needs evidence before it counts as done.
  • Send shows the exact agents and message first. Nothing goes out until you click.
  • Receipts record every send: queued, failed or uncertain. An uncertain send is never retried for you.
The Approvals view on a phone, in the Fantasy design: 4 waiting for you, 2 notices, with filter chips by kind. First a waiting chat, Planning in Docs Site, asks whether the release should go out today, with a Reply button under the question; then a decision in Checkout API about signed cookies.
Approvals on a phone, in the demo city: what waits for you comes first, and a waiting chat is answered from its row.
The training grounds: a fenced yard where idle agents spar, practise archery and rest by a campfire.
When nothing needs them, idle agents drill in the Fantasy training grounds.

See what they waste.

A prompt cache that expires while a chat sits idle is paid for again on its next turn. Co-ordinate counts every reload from the transcripts on your machine, and says why it happened.

  • A cache countdown on every chat: warm, cooling, or cold with what a reload would cost.
  • The waste ledger: money lost to reloads in the last 7 days, by chat and by reason.
  • Startup size: what a new chat loads before it starts, and when a fresh start would pay back.
  • Model fit finds big-model turns a smaller model could have done; Outcomes show what each task, pull request and commit cost.
The Waste panel on the Board, Stone Age design: $15.42 avoidable in the last 7 days, 23 reloads a warm cache would have saved, split by reason (cache expired 14, model switched 4, prefix rebuilt 3, compaction 2) and by chat, with Planning in Docs Site first at $4.86.
The Board's Waste panel in the Stone Age design. The demo city's numbers are made up.

Capped runs stop on their own.

A capped run starts Claude Code or Codex on one prompt, again and again, until one of its four caps stops it: iterations, dollars, tokens or minutes. There is no unlimited run.

The Capped runs panel on the Board, Lunar design: Fix flaky tests in Checkout API running at 7 of 20 iterations and $1.84 of $5.00, with a Kill button; Docs link check, a Codex run, at 3 of 10 iterations; Lint until clean paused after the same error three times, with Resume and Stop.
The Board's Capped runs panel in the Lunar design: two runs under their caps, one paused by the stuck detector. The demo city's numbers are made up.
Caps first
The start button shows the dollar cap (a Codex run, which has no prices, stops at its token cap). At any cap the run stops at once, and the kill switch is on the Board and on its building.
Stop conditions
Tests green, a small diff, no progress, a pull request merged.
Stuck detector
A run that repeats itself is paused, and waits in Approvals for Resume or Stop.
Repeat prompts
What your own /loop, wake-ups and scheduled prompts cost, and what the next 7 days will.
The Wasteland design: a settlement of repaired workshops and towers on an island, with a salvage yard in the middle.
Wasteland design: idle agents work the salvage yard in the middle.

Free to watch. It stays on your machine.

Co-ordinate reads files already on your machine. Watching, discovery and imports make no model calls. Tokens are spent only when you click Send, Train, Build HQ, or start a capped run or a bench run, after it shows what it will do. Adding a folder as an HQ is free.

Model calls to watch
0
Listens on
127.0.0.1, your machine only, behind a key that only your account can read
Leaves your machine
Nothing on its own. What can, such as a send or the team export, goes only on your click; the app's Privacy page lists each one.
Never reads
Credentials, such as ~/.codex/auth.json
Claude + Codex
Same buildings, separate allowance meters, with a Claude subscription or an API key. A missing reading says unavailable, never zero. Cursor and Gemini CLI chats show too, read only.
Server
Python 3.10 standard library, no packages
Telemetry
None. No analytics, no update check
Runs on
Windows today. macOS and Linux run the server and the page; opening an exact chat in a terminal is Windows only.

Light enough to leave open all day.

It sits beside your agents from morning to night, so it is built to stay out of the way.

Measured in Chrome on a quiet laptop on 9 October 2026, over several runs, with test cities of 1, 10 and 16 projects in all six designs.
Window focusedAbout 30 frames a second
Window in the backgroundOne still frame, then nothing
Tab hiddenNo drawing at all
Page script4.7 to 8.0 % of one CPU core in the newest run; up to 11 % in other runs that day
City size1 to 16+ projects in 44 to 88 draw calls, under a budget of 140
The Edo design at night: a castle town on an island under a snow-capped volcano, project signs glowing.
Edo at night. Time of day and weather are yours to set.
The temple garden from above: raked gravel around three stones, with agents kneeling on the veranda.
The temple garden. Rake it while you wait.

Also in it.

Built and tested; all of it arrives with the first public release, 1.0.0.

Around the city

  • Activity: notifications, requests to agents, tasks, a feed, commits and a daily digest.
  • Ctrl+K search, keyboard shortcuts and a phone layout.
  • Build HQ with the tool you pick, or add one folder as an HQ for no tokens.
  • Spending caps per project, and guardrail hooks against secrets, destructive commands and forbidden paths.

Skills and designs

  • 21 knowledge packs, each with its cost in tokens, installed on a click.
  • Training from reviewed lessons, and an A/B bench for your CLAUDE.md and skills.
  • Six city designs, each with its own yard and calm place, or a design of your own.

Coming soon to GitHub.

Free and open source (MIT) for one person on one machine. It needs Python 3.10 or newer and no other packages. Join the waitlist and you get one email when it is out.

One email when Co-ordinate is out, and first a link to confirm your address. Your address goes only to the mailing-list service that sends it, and every email has an unsubscribe link.

  • Zero tokens to watch. Watching makes no model calls.
  • Nothing leaves on its own. What can, leaves only on your click.
  • It never reads credentials. Keys, tokens and auth files are off limits.
  • It spends only on your click, after showing what it will do.