!v1.0.9 · New Workbench Canvas →

Hand AI what you want done.
It finishes and tells you.

Say what you want done and AI gets to work. Pick “Until done” and it breaks the job into steps and keeps going. When something needs your call, its card lights up and waits for your sign-off. No coding, no babysitting.

Download for Windows macOS (Apple Silicon)

Also for Intel Macs · Linux · Downloaded 8,607+ times

SpectrAI · Workbench Canvas

Workbench

FitTidyNeeds meToday's timeline

2Awaiting you

Focus: V2EX watchEsc

Intel watch → Inbox

SpectrAI dev

Learning

linuxdo intel watch

Idlenext in 1 min

last 09:53 · 25 logged

52pojie hot posts

Runningpage 2

last 07:36 · 6 pushed

V2EX hot posts

Idlenext in 57 min

last 07:49 · 8 pushed

AI news digest

Idlenext 20:30

AM 08:30 · 271 in, 12 picked

Inbox

173today

SpectrAI positioning talk

Running3m44s

drafting canvas prototype

SDK rewrite · team

Running2/6 running

sidebar filter + state machine

Release check

Needs you9 min ago

installer built, awaiting sign-off

Broadcast queue not consuming

Runningjust now

reading logs

Founder skills track

Needs you1 h ago

lesson 3 passed, ready for 4

Note: tidy reading notes

Make sessionMake goal

Needs you · broadcast queue

Cause found: consumers do not resubscribe after restart. Pick a fix:

A Resubscribe on reconnect

B Add a dead-letter queue

Say what you want done, press Enter…

Just do itUntil doneTeam upSchedule

Today's timeline · scrub back through today’s scheduled runs Group summaries · decisions and next steps, straight into your queue Notes · jot it down now, dispatch it as a session or goal later

AI can do more and more.
Keeping track is the hard part.

Once you have several sessions, goals and scheduled jobs running at once, the problem is no longer starting them — it is seeing what they are doing. That is what the Workbench is for.

BeforeWith the Workbench

Too many sessions to followA long sidebar list — no idea which is running, stuck or long finished.

Sessions, goals, scheduled jobs and teams become cards, grouped by project, status at a glance.

AI is waiting and you never noticedIt has been waiting for your reply or sign-off for hours.

Cards that need you light up and get stamped; “Awaiting you” lists them all, failures included.

Gone quiet mid-runIs it stuck or still thinking? You cannot tell.

Anything running with no activity for 10 minutes gets an amber hint and moves to the front.

Who feeds whom?Watches and sub-tasks are connected, but you cannot see how.

Links show whose output feeds whom; right-click “Focus this line” and the rest fades.

A busy day, but what got done?Reviewing means opening everything one by one.

Today’s timeline replays every scheduled run; group summaries pull out decisions and next steps, straight into your queue.

Say it in one line. Pick how it runs.

Type what you want at the bottom of the Workbench, then pick how it should run. Here are a few lines you can copy as-is.

Just do it

Ask like you would in chat

AI starts right away and reports back when done.

Look at recent changes in this project, find the one issue most worth fixing, fix it, then run the tests

Until done

Turn it into a goal

AI breaks it into steps and keeps pushing, round after round, until it is truly done.

Fill in the unit tests: list uncovered modules, add tests one by one, until everything passes

Team upPro

Send in an AI squad

They split the work and review each other; anything below the bar goes back for a redo.

Refactor the login module into its own service without changing the API, then cross-review

SchedulePro

Run on a schedule

Runs daily or weekly and brings the results to you.

Every day at 9am, check the project builds and dependencies have no security alerts; summarize any problems for me

Not ready yet? Press Alt + Enter to save it as a note — nothing runs until you dispatch it.

From start to sign-off,
every step is checked.

Work on several things side by side

Workgrid lays several sessions out on one screen. Drag to split — no more window switching.

Drag to split; open as many sessions side by side as you like
Split ratios are remembered and restored
Code, monitoring and docs — arrange them however you like

Workgrid3 sessionsratio saved

Refactor loginRunning

Split it out, keep the API

OK. Listing callers first, then migrating one by one, then running tests.

Debug Nginx 502Needs you

Upstream timeout found — two fixes, your call.

Weekly reportDone

Experts review it — and send it back if it falls short

Bring in an architect, a security officer or a QA expert from the Expert Store — or seat the whole council. Every result is scored on five dimensions; one veto sends it back.

Completeness, robustness, security, code smell, architecture contract
Below the bar goes back to the lead agent until it passes

Teams & experts: Pro and up

Cross-reviewLogin refactor · round 2

AArchitect · contract8.6Pass

SSecurity · safety5.2Veto

QQA · robustness8.1Pass

RReviewer · code smell7.9Pass

PProduct · completeness8.4Pass

!Security vetoed: the password-reset endpoint has no rate limit. Sent back for a redo; it will be reviewed again.

Review AI changes before they touch your main branch

Each task works in its own isolated copy, away from your main branch. Like the diff? Merge in one click. Don’t? Throw it away.

File-by-file diffs, every line accounted for
A silent pre-check flags conflicts before you merge
Parallel tasks each get their own copy

Isolated copyworktree/fix-reset3 files changed

auth/reset.tsauth/rate-limit.tsreset.test.ts

export async function resetPassword(req) {

+ await rateLimit(req.ip, { max: 5, perMinutes: 10 })

const user = await findUser(req.email)

- if (!user) throw new Error('not found')

+ if (!user) return ok() // don't leak which accounts exist

await sendResetMail(user)

}

Pre-check passed · no conflictsDiscardMerge

Turn routine work into a pipeline

Draw the process once: AI picks the branch, failures retry with the error attached, and nothing moves on until tests pass.

Branches, retry loops, fan-out and merge
Build and tests are gates — fail and it goes back

Workflows: Ultra

Built-in browser · open web pages in the app Project library · no more re-explaining context Command center · token spend and risks on one page

Away from your desk?
Still your call.

AI keeps running on your computer; decisions come to your phone. Tap to reply, sign off or pick an option, and it carries on.

Mobile appTelegramFeishuWeCom

Remote channels depend on your plan: Telegram from Pro; Feishu, WeCom and full mobile control on Ultra. Set up mobile →

9:41

Workbench

3 running · 1 needs you

! Broadcast queue not consuming

Needs you two fixes, your call

A: resubscribeB

AI news digest

Running 271 collected

Weekly report

Done 3 min ago

31+ model services. Use the right one for each job.

Official CLIs, major models, local models and relays all plug in. Send each job to the model that suits it.

Official CLIs

Claude Code
Codex CLI

Major models

Claude · GPT · Gemini
DeepSeek · Qwen
GLM · MiniMax

Local

Ollama
vLLM
LM Studio

Relays & compatible

OneAPI · NewAPI
OpenAI-compatible

Start free. Upgrade when you need to.

The free plan works out of the box. Upgrade for teams, schedules and remote sign-off. Activate in the app with a code.

Free

Hand AI what you want done — it works it through until done. Ready out of the box.

Goals, Workbench, Awaiting your decision & Xiaoguang
AI does it directly · until done (single runner)
Basic workgrid split-screen
Basic prompt optimizer
No team-up, Xiaoguang with experts, or scheduled goals
No fixed workflows
No mobile & IM remote control

Download free

Personal powerhouse

Pro

High-performance AI workbench for professional developers.

Team up: AI Team collaboration & cross-review
Xiaoguang with experts
Sessions can autonomously dispatch sub-tasks
Scheduled goals
Dynamic persistent memory & project knowledge base
Telegram remote channel & cloud relay
No WeCom/Feishu full-channel support

Get a code

Best for automation

Ultra

24/7 automated, unattended engineering hub.

Fixed workflows (workflow orchestration) & prototype design mode
Feishu + WeCom + mobile full-channel control
More scheduled goals & custom event hooks
Goal drift detection & auto-reject
Model service performance tracking & quality feedback loop
Priority technical support & dedicated channel

Get a code

01Buy a code

Choose a plan on the store page and you will get a code after payment.

02Activate in the app

Update to the latest version, then Settings → Account/Plan → Redeem code.

03Questions?

Buying and renewing: FAQ; refunds: refund policy, or ask in the group.

Hand AI what you want done.It finishes and tells you.

AI can do more and more.Keeping track is the hard part.