The operating layer for AI work

Give AI a task.
Get finished work.

Plaux turns your Mac into a visible, controllable team of AI operators. It plans, uses approved apps, browser sessions, files and terminal tools, then leaves a checked result, reusable skill or UI map behind.

Open your Plaux cabinet Explore capabilities

Works in your apps You see every step Data stays on your Mac

plaux / live run running

PLAUX TASKCompare 20 competitors

01:42

01
Find the companiesWebsites + social media
✓
02
Collect prices and offersBrowser + spreadsheet
✓
03
Remove duplicatesData validation
04
Save the comparisonReady for the team

⌘Applicationsworking

✦Mac controlvisible

✓Resultchecked

Drag the sceneLive execution model · illustrative

One system, many work surfaces

Everything needed to move from request to reliable result.

Plaux keeps the operator, its memory, evidence and permissions together instead of scattering a job across tabs, chats and scripts.

◎

Operators and routing

Assign a task to the right model, skill and project chat. Fail over within the provider, then across providers when the task needs it.

Claude · Codex · Gemini · Grok · Kimi · OpenRouter

⌁

Memory that earns trust

Keep verified routes, application knowledge, UI maps and work graphs. Reuse only what has evidence and see what changed.

Routes · nodes · links · traces · reliability

▣

Real Mac execution

Work in visible apps, authorised browser sessions, files and terminal tools. Every run has a mode, budget, stop control and receipt.

Accessibility · browser · files · capture

◌

Capture & visual evidence

Collect one screenshot or a gallery in chat, inspect frames, attach the evidence to a report and keep source material local by default.

App frames · websites · galleries · reports

↗

Cloud workspace, by consent

Sync encrypted projects, chats, skills, UI maps and graphs to your cabinet. Community stays local; Pro can publish selected structure.

Explicit opt-in · encrypted payloads · revocable devices

✓

Outcome-driven work

Set the goal, time window, repeat policy and responsible project/chat. Plaux reports the result, failures, cost and next action.

Goals · schedules · budgets · verification

Write the task◆Plaux works in your apps◆Watch the progress◆Check the result◆

Your AI, one workspace

Use the AI services you already have.
Plaux keeps the work together.

Connect Claude, Codex, Gemini, Kimi, Grok or open models. Plaux sends each part of the job to a suitable model, while your task, permissions and result stay in one place.

PLAUX / OPERATOR BUS 6 provider paths

One goalPlaux coordinates

Shared context · explicit access · visible result

AnthropicClaude Code

CLI

Mac + terminal

Fable 5
Opus 4.8
Sonnet 5
Haiku 4.5

◎

OpenAICodex CLI

CLI

Code + terminal

GPT-5.6 Sol
Terra
Luna
GPT-5.5 / 5.4

✦

GoogleGemini CLI

API

Planning + terminal

Gemini Auto
1M context

◒

Moonshot AIKimi Code CLI

CLI

Code worker

Kimi for Coding
262K context

⌁

OpenRouterOpen models

API

Your selected routes

Open-source catalogue
Model picker

𝕏

xAIGrok

API

Reasoning + terminal

Selected Grok models
Model picker

One service is enough to start. If you connect several, Plaux can choose a better fit for a specific step. It never gives a provider access unless you allow it.

Three simple steps

You ask.
Plaux does. You check.

It is more than a chat: Plaux can actually use the programs on your Mac and bring back a result you can inspect.

1. Write the task

Explain the result in ordinary words. Plaux keeps the request in chat so it is always clear who started the work and why.

One clear request
Your files and instructions
Only the access you allow

2. Watch it work

Plaux opens apps, types, searches, sorts files or uses the terminal. You see the progress and can pause or stop it.

Visible actions
Time and progress
Pause or stop

3. Check the result

Plaux saves what was done and the proof it produced. A successful approach can help similar work run better next time.

Files, screenshots and logs
Result you can review
Useful steps remembered

Spend less on complex work

Use expensive AI
only where it matters.

For a large task, Plaux can ask a strong model to make the difficult decisions and give the simpler steps to faster, cheaper models. A separate check catches mistakes before they become costly.

01 Split the job only when it saves time or money 02 Give each helper only the information it needs 03 Count the cost only after the result is checked

PLAUX / GOAL TREE BOUNDED

OUTCOMEPublish a verified releaseBudget · permissions · evidence

◆

FRONTIER PLANNERResolve ambiguity

Decompose, choose boundaries and define acceptance.

judgment

WORKER ABuildbounded context

WORKER BInspectbounded context

WORKER CDocumentbounded context

↔RECONCILERresolve overlap ✓REVIEWindependent lens ■SUPERVISORverify or stop

Adaptive topologySolo remains valid when coordination would cost more than the split saves.

Automatic model choice

A suitable model
for each part of the job.

Plaux looks at how difficult and risky each step is. It can use a faster model for routine work and a stronger one for a hard decision. Another AI service is used only if you connected it and the task really needs it.

PLAUX / MODEL SWITCHBOARD POLICY ACTIVE

INCOMING CONTRACTShip a verified product change4 work steps · one acceptance boundary

ROUTING

HIGH AMBIGUITYArchitecture decision

FRONTIERClaude Codesame provider

LOCAL ROUTE

BOUNDED CODEGenerate focused tests

FASTClaude Codetier switch

AUTOMATIC

VISIBLE APP STATEVerify the interface

GUI CAPABLEClaude + Handsprovider permission

TRUST GATE

PROVIDER-SPECIFIC CHECKAudit the final patch

BALANCEDCodexsecond provider

TASK REQUIRED

TASK→CAPABILITY→TRUST→MODEL→EVIDENCE

SAME PROVIDERAutomatic tier selection CROSS PROVIDEROnly when the task requires it AUDIT TRAILEvery switch remains visible

How routing changes The goal stays fixed. Each step receives only the context, model strength and provider access it needs. A lower price alone never justifies switching providers.

Capture Studio

Turn work into content.
Keep it as evidence.

Capture an area, an app window or a short clip without leaving the task. Plaux records the capture in chat, keeps the media local and lets you reuse it in a post or a verifiable report.

ContentCrop for 16:9, 1:1 or 9:16. Attach it to chat or prepare it for X, Telegram and Shorts.

EvidenceKeep the app, device time, task and result together in a local project report.

⌾Local by default. The model can use selected frames, text and events — not the entire video.

PLAUX / CAPTURELOCAL ONLY

Selected application

16:9

00:24 / 00:30

⌖ Cursor◖ System audio⌁ Sensitive fields

Examples from everyday work

Give Plaux a job
you would give an assistant.

Ask it to collect information, update a table, work with social media, prepare a server or clean up documents. Plaux uses the same apps you use, while you can watch or stop the work.

Visible taskPlaux · working

⌕

YOUR REQUEST

Build a clean map of relevant Telegram channels

01
Open Telegram DesktopUse the signed-in app already on this Mac
✓
02
Search topic by topicHuman-paced navigation through the native interface
✓
03
Check relevance and duplicatesKeep only public, useful matches
✓
04
Write verified rows immediatelyThe working document stays up to date
✓

RESULTA ready-to-use market map instead of hours of manual searching.verified

RUN → ROUTE Plaux learns only after the result is verified.

Replay economics

Reason once.
Replay what was proven.

A model-only desktop agent repeatedly looks at the screen and reasons about every click. After Plaux verifies a successful route, deterministic steps can be replayed while the model returns only for changed state or a new decision.

In one tested desktop workflow, the builder observed roughly 100× lower cost than a model-only run. This is a single workload, not a universal guarantee.

MODEL-ONLYMore loops

See→Think→Click→Repeat

VERIFIED REPLAYFewer fresh decisions

Recall→Run→Check

Tested in real work

What Plaux has already done in real workflows.

Builder- and client-reported outcomes show what was tested, not what every user should expect.

20K+

Topic-relevant Telegram entities catalogued

A client used the resulting dataset for competitive targeting and reported approximately $38K in additional revenue. Not independently verified.

Market research · client-reported

A social account kept moving

Plaux navigated X, joined relevant conversations and gained 8 organic followers in one unattended experiment.

Social operations · results vary

1 run

Server and domain taken live

Infrastructure, TLS, domain routing and the product website were configured through one visible task flow.

Launch operations · verified

Messy → usable

A sales document finally became workable

A large document the team struggled to systemize was cleaned, sorted and turned into a usable operating structure.

Back office · verified workflow

◎ Plaux works through standard app interfaces where APIs are missing. Operators remain responsible for service rules, permissions and the work they authorize.

Memory that learns from work

Plaux remembers what worked — not everything you wrote.

After a checked result, Plaux remembers the useful steps. Failed approaches are marked as unreliable, and old ones lose priority. When a similar task appears, Plaux can reuse only the proven part without copying your private content.

STRENGTHENVerified edges gain confidencepromote
QUARANTINEFailed branches stop steeringisolate
DECAYStale routes wait for new proofrefresh

The visual explains graph behavior. Values are illustrative, not real-world performance claims.

EXPERIENCE MODELLOCAL · CONTENT EXCLUDED

NEXT TASK RECALLS7-node proven subgraphROUTE CONFIDENCE92%illustrative

Orchestration Graph

Remember how the agent team worked.

A dedicated Memory view compares compatible multi-agent runs: how the goal was split, which role consumed context, where work overlapped, what the review caught and whether the final outcome was verified.

CONTEXTPlanner and worker load stay separate
COORDINATIONDuplicate branches and reconciliations are visible
ECONOMICSCost is tied to a verified result

RUN COMPARISONcompatible cohort only

Goal decompositionstable

Context isolationmeasured

Duplicate workwatch

Verified outcomerequired

Built in the open

Choose the workspace that matches your work.

There is no public sign-up. Access is granted from the Plaux cabinet; invited users receive a one-time email code and only the edition they were given.

⌘

OPEN SOURCE

Community

Local

A small, inspectable local edition for learning the operator model and running useful work on one Mac.

✓ Local projects, chats and files
✓ Local skills, traces and memory
✓ Basic operators and UI maps
× No cloud sync or private releases

Ask for Community access ↗ Built to be useful locally, with a clear path to Pro.

✦

PROFESSIONAL

Pro

Invite

For people who want cloud continuity, operator memory and a full production workflow without the developer-only surface.

✓ Encrypted cloud projects and chats
✓ Skills, UI maps and graphs across Macs
✓ Browser, capture and outcome workflows
✓ Signed updates and revocable device access

Independent developer

Built by Ilia Volkov.
For people who finish the work.

Plaux began with a practical problem: even capable models lose time, context and money when they operate real software. Ilia is building the missing execution layer — local, inspectable and shaped by real work rather than demos.

DIRECT CONTACTChoose the right channel

Replies personally

COMMIT 01Own your contextLocal-first by default

COMMIT 02Verify the outcomeEvidence over confidence

COMMIT 03Learn from the runRepeats can increase token savings

Release contract

A feature is not shipped
because it exists.

01 · BUILDImplemented locally

The feature works in a development build. This alone does not make it a public release.

02 · PROVETested and documented

Behavior, permissions, migration and rollback are checked against real evidence.

03 · PUBLISHReleased deliberately

A signed update reaches users only after explicit approval and a verified recovery path.

Questions in plain language

What Plaux is and what it does.

Short answers without technical terms. If you can describe a task to another person, you can describe it to Plaux.

What is Plaux?

Plaux is a program for Mac that lets AI do real work, not only answer questions. You write a task in chat. Plaux can open apps, use files and the terminal, complete the steps and show you the result. Everything stays in one place: the request, permissions, progress and proof of what was done.

Why build an ecosystem instead of one universal agent?

Different AI models are good at different things. One may plan better, another may write code faster, and a simple tool may be enough for a repeated click. Plaux connects these parts so you do not have to move the task between several chats and programs yourself.

How does Plaux work across Mac applications?

Every job starts with a message in chat. Plaux can click and type in visible Mac apps, work with project files, use the terminal and start scheduled tasks. You can watch, pause or stop the work and ask Plaux to wait for approval before important actions.

Does Plaux upload my screen or project data?

No, not by default. Projects, screenshots, videos, work history and learned steps stay on your Mac. Cloud sync or sharing can happen only as a separate feature that you choose to turn on.

How can Plaux reduce the cost of AI automation?

Plaux can use an expensive model only for the difficult decision and cheaper models or ordinary tools for the simple steps. If a successful sequence can be safely repeated, the AI does not need to think through every click again. Similar repeated work gives more chances to save tokens, but savings are not guaranteed every time.

How do orchestration and sub-agents fit into the ecosystem?

For a large job, Plaux can divide the work between several AI helpers. One makes the plan, others handle clear pieces, and another checks the result. Plaux uses this only when dividing the work is actually useful; a small task can stay with one agent.

What does Plaux remember between runs?

Plaux remembers the steps that led to a checked result. It does not treat every sentence in chat as a fact. Good approaches become easier to reuse, failed ones are avoided, and private task content stays on your Mac.

Which parts of Plaux will be open source?

The Community edition is built around an open core and is being prepared for a public source release. Pro adds cloud continuity, shared workflows and deeper controls.

Plaux Community