Every model. Every subscription. One secure harness.

A complete coding harness with its own engine: guarded tools, approvals, a browser, terminal, diffs and memory, for any model you bring.

macOS app · Coming soon See how it works

Usage page in Settings

Build a Usage page in Settings with this week's spend per day.

Look at the usage data
Split the work across three helpers
Draw the weekly spend chart
Add the page to Settings

Ran 3 agents

Spend per day

This week, all providers

$20$15$10$5$0

Edit files

Allow Crewroom Agent to edit the following file?

web/src/settings/routes.ts

{ path: "appearance", page: Appearance },

+ { path: "usage", page: UsagePage },

{ path: "models", page: Models },

Always allow DenyEscAllow once⏎

Edited routes.tsnow

Done. Settings has a Usage page with this week's spend per day. All 14 settings tests pass.

Reply to continue

Auto GPT-5.5 Medium

Details Changes Terminal

Subagents

AdamSonnet 4.6

RicardoGPT-5.5

KDeepSeek V4.1

Changes

UsageChart.tsx+31 −4

usage.ts+7

Uncommitted changes+38 −4

web/src/settings/UsageChart.tsx+31 −4

@@ -1,8 +1,35 @@

1 import { ShowCard } from "../cards";

2-export const UsageChart = () => null;

2+export function UsageChart({ days }: Props) {

3+ return (

4+ <ShowCard type="bar" title="Spend per day"

5+ labels={days.map((d) => d.label)}

6+ unit="$" highlight={today()} />

7+ );

8+}

web/src/usage.ts+7

@@ -12,3 +12,10 @@

13+export const spendByDay = (rows: Row[]) =>

14+ groupBy(rows, (r) => r.day).map(sum);

9:41

Usage page in Settingscrewroom-web

+38−4

Build a Usage page in Settings with this week's spend per day.

Thought for 4s

Edited 2 filesUsageChart.tsx, usage.ts

+38−4

Editedroutes.ts

Done. Settings has a Usage page with this week's spend per day.

Working· 0:12

Ask anything

Crewroom Agent

Edit this file?

web/src/settings/routes.ts

ReasonAdd the Usage page to Settings.

DenyAllow once

Always allow

01 · Planning

Plans before it builds.

Crewroom writes a plan you can read and approve, then ticks it off as the work lands.

Learn about planning →

Weekly digest email

New task

Add a weekly digest email for project owners.

Reading scheduler.ts

Here's the plan. I'll start once you approve it.

Edited scheduler.tsnow

Created digest.tsnow

Edited NotificationsPage.tsxnow

Ran a command6 pass

Done. Owners get the digest every Monday at 08:00, and can turn it off in Settings.

Add the digest job to the scheduler
Build the email from the week's sessions
Add an on/off switch in Settings
Test the job with a fake clock

Ready to code?

Weekly digest email

Add a digest job to the scheduler
Build the email from the week's sessions
Add an on/off switch in Settings
Test the job with a fake clock

Copy Save as .md Build it in a new chat

Tell Crewroom Agent what to change No, keep planning Yes, manually approve edits Yes, auto-accept edits

Reply to continue

Plan first DeepSeek V4.1 High

Details plan.md

Weekly digest email

A short Monday email to each project owner.

Why

One place to see what the crew did.

Steps

Add a digest job, Mondays at 08:00
Build the email from the week's sessions
Add an on/off switch in Notifications
Test the job with a fake clock

Out of scope

Daily digests
Slack delivery

Spend dashboard

Spend this month

Make me a dashboard of this month's AI spend and budget per project.

Reading usage data

Here's your dashboard. It refreshes from your usage data.

Spend this month

Dashboard · View anytime from the sidebar

Preview

Reply to continue

Auto Gemini 3.1 Pro Medium

Budget left

$386

of $600 · 64% left

Top model

GPT-5.5

$92 this month

Budget left by project

crewroom-web52% left

api-server20% left

docs-site71% left

mobile88% left

Edit

Spend per day

November, by provider

AnthropicOpenAIDeepSeekGemini

Edit widget

Color

02 · Dashboards

Dashboards your crew builds.

Ask for a view of anything: spend, budgets, progress. You get a live dashboard that stays in your sidebar.

See dashboards →

03 · Cost view

Deploy the docs site

Deploy the docs site to Vercel

Ran npm run build42 pages

The build is ready. Deploying needs a Vercel token, so I'll ask you for it.

SecretVERCEL_TOKENPassed to the command and kept private2:14 PM

Ran vercel deploy --prodlive

Deployed.

Model saw: the command's output and exit code. Never the value.

Secret

Pass VERCEL_TOKEN to this command?

Reason: “Needed to deploy the docs site to Vercel”

~/Projects/docs-site

vercel deploy --prod

Paste the secret

Goes only to this command while it runs. Never to the chat, the model or a file.

DeclineEscPass to command⏎

Reply to continue

Kimi K2.5 High

04 · Secrets & safety

Type a secret. The model never sees it.

When an agent needs a key, you type it into Crewroom's own window. It goes straight to that one command, never to the chat or the model.

See how secrets work →

05 · Phone app

Your crew, in your pocket.

Pair your iPhone with a QR code. Approve, steer and start tasks from anywhere, straight to your own Mac with no company server in between.

Phone app · coming soon

Fix the flaky login test

Read login.test.ts

Edited login.test.ts

Ran npm test · 48 passed

Paired with MacBook Pro · Wi-Fi

9:41

Fix the flaky login testapi-server

+12−3

Fix the flaky login test

Readlogin.test.ts

Editedlogin.test.ts

+12−3

Ran npm test48 passed

Working

Queued · 1

also update the changelogSteer now

Ask anything

Crewroom Agent

Run this command?

npm test

ReasonCheck the login test passes.

DenyAllow once

Allow asks for Face ID

Schedules

ScheduleNext run

06 · Bots & schedules

A crew that works while you sleep.

Give a bot a role and a schedule. It runs on its own, wakes when CI fails, and paints in its second eye when it's done.

Meet the bots →

07 · Memory

Every project remembers on its own.

Each project keeps its own memory. Agents save what they learn; you can review, edit or forget any note.

See how memory works →

api-server

Add a health check route

Edited src/routes/health.ts

Added GET /health. It returns 200 when the server is up.

Run the tests

Ran bun testfailed

They need TEST_DB set. I ran them with bun, since this repo does not use npm.

Reply to continue

Memory saved

Memoryapi-server

TimelineRecently learnedMissedSearch memoriesAdd

What agents and capture rules saved lately, newest first. Agents use these right away; Forget takes one out of recall.

LearningRun tests with bun test; TEST_DB must be set14:32

Forget

FactAPI routes live in src/routes09:41

Forget

DecisionErrors return JSON problem detailsTue

Forget

MistakeUnverified claimNever run migrations on the shared databaseMon

Forget

DecisionDeploys go through Vercel11:20

Forget

FactPages live in content/docsTue

Forget

LearningThe link checker runs in CIMon

Forget

Memory lives in Crewroom's local database, one per project.

How it works

One task, end to end.

Follow one task through the app, the daemon, the engine, the guard and your model. The messages on the right use the real protocol.

One taskFix the flaky login test in api-server

System graph2 parts sit outside your Mac

Your Mac

Desktop app

chats · approvals

Tauri · window

Phone

follow · approve

iOS · your tunnel

Crewroom daemon

routes every chat

Bun · 127.0.0.1

Memory

per project

Chat log

JSONL

Crewroom Agent engine

one per chat · pipe fd 3

Bun · child process

Model provider

claude-sonnet-5-5

Shell guardseatbelt

Tools

bash · edit · read

MCP servers

stdio · per project

You send the task. The daemon queues it.

The app hands it to the daemon on 127.0.0.1, which keeps one engine per chat.

You send the task. The daemon queues it for this chat's engine.
The engine starts the turn and calls your model. It gets the key from the daemon, then streams from your provider.
The model asks to run a command. The guard checks it against your rules.
A risky step waits for you. Approve it in the app or on your phone.
The tool runs inside the guard. The output goes back to the model.
The model finishes. The reply, the cost and the memory are saved.
Only model calls leave your Mac. Your phone connects through your own tunnel.

daemon ⇄ engine · fd 3JSON lines

1→{"id": 7, "method": "turn/start",

2 "params": {"input": [{"type": "text",

3 "text": "Fix the flaky login test…"}]}}

4←{"id": 7, "result": {"turn": {"id": "t_41",

5 "status": "inProgress"}}}

6←{"method": "turn/started", "params": {

7 "turn": {"id": "t_41", "status": "inProgress"}}}

8←{"id": "e3", "method": "credentials/get",

9 "params": {"model": "claude-sonnet-5-5"}}

10→{"id": "e3", "result": {"apiKey": "sk-ant-•••"}}

11←{"method": "item/started", "params": {

12 "itemId": "i_2", "kind": "commandExecution"}}

13←{"id": "e4", "method":

14 "item/commandExecution/requestApproval",

15 "params": {"turnId": "t_41", "card": {…}}}

16→{"id": "e4", "result": {"decision": "accept"}}

17←{"method": "serverRequest/resolved",

18 "params": {"requestId": "e4", "reason": "owner"}}

19←{"method": "item/commandExecution/outputDelta",

20 "params": {"itemId": "i_2", "delta": "24 pass"}}

21←{"method": "item/completed",

22 "params": {"itemId": "i_2"}}

23←{"method": "item/started", "params": {

24 "itemId": "i_3", "kind": "agentMessage"}}

25←{"method": "turn/completed", "params": {

26 "turn": {"id": "t_41", "status": "completed",

27 "durationMs": 9180}}}

Real message names, fields trimmed. → daemon to engine, ← engine to daemon.

Execution traceturn t_41 · illustrative timing

02 s4 s6 s8 s

turn/start → completed

credentials/get

model · round 1

guard · rules

waiting on you

bun test · seatbelt

model · round 2

Working on itThe reply appears when the turn completes.

+6 −2·login.test.ts·$0.14

08 · Every model

Bring the models you already pay for.

Keep going on the next model…

Same chat, same context

ChatGPTYour plan
GitHub CopilotYour plan
GrokYour plan
Google GeminiYour plan
ClineYour plan
AnthropicAPI key
OpenAIAPI key
DeepSeekAPI key

KimiAPI key
Z.AI / GLMAPI key
QwenAPI key
MiniMaxAPI key
OpenCode GoAPI key
OpenRouterAPI key
OllamaLocal · free
LM StudioLocal · free

FAQ

Questions, answered.

The short version of how Crewroom handles your models, keys and code.

Start your crew tonight.

Bring the models you already pay for. Everything runs on your Mac.

macOS app · Coming soon Phone app · Coming soon

Every model. Every subscription. One secure harness.

Plans before it builds.

Dashboards your crew builds.

See where every token goes.

Type a secret. The model never sees it.

Your crew, in your pocket.

A crew that works while you sleep.

Every project remembers on its own.

One task, end to end.

Bring the models you already pay for.

Questions, answered.

What is Crewroom?

Which models can I use?

Can I use my Claude Pro or Max subscription?

Do local models work?

Can I switch models in the middle of a chat?

Does the model ever see my secrets?

What can an agent change on my Mac?

Where are my API keys kept?

How does the phone app connect?

Does Crewroom send my code anywhere?

Is there a Windows version?

How much does it cost?

Start your crew tonight.