PixLab released Think-Act for Windows on September 18, 2026. It is a desktop app that puts an AI task agent, a writing assistant, OCR, a screen recorder, reusable local Skills, and local MCP connectors in one place. Supported actions can be set to ALLOW, ASK, or BLOCK. It runs on 64-bit Windows 10 and Windows 11, needs a PixLab account, and is free to install and try. PixLab's downloads page puts the macOS release in Q1 2027.
That makes it a different kind of product from a self-hosted agent like OpenClaw or Hermes. Think-Act is for one person at their own PC who starts a task and reviews the result. OpenClaw and Hermes are services you run so an agent can answer in chat, run on a schedule, and keep working when nobody is at the keyboard.
Sources for the Think-Act side are PixLab's launch post Think-Act Desktop Agent Is Now Publicly Available for Download and the product pages for the Desktop Agent, integrations, security, downloads, and the FAQ, all checked on October 10, 2026. The OpenClaw and Hermes side reuses our OpenClaw vs Hermes Agent comparison.
The short answer
Use Think-Act when the work starts on one person's Windows desktop: documents, scanned PDFs, screenshots, selected text, or a recorded process, and that person is there to approve actions and check the output. Use OpenClaw or Hermes when the agent has to run as a service on a host you administer, serve more than one person, reply in messaging channels, or run on a schedule. If you need typed access to a business system, an MCP server you review can serve either one.
What Think-Act includes
- A Desktop Agent for bounded multi-step tasks. You describe the job, attach the files or context it needs, and optionally add a success goal, which is a checkable finish condition such as "every invoice has a total extracted."
- A Writing Assistant that works across Windows apps such as Word and Excel through a global shortcut (Alt+Space by default). It rewrites, summarizes, translates, changes tone, and drafts.
- OCR and document extraction for screenshots, scans, images, and image-based PDFs, with the output usable as context for the next step.
- A Screen Recorder whose recordings can become transcripts, how-to documents, bug reports, or the start of a reusable Skill.
- Local Skills, which save instructions and checks for work you repeat.
- Local MCP servers as connectors. Node.js or Python is needed only when a specific MCP server requires it.
PixLab's FAQ says the Desktop Agent runs when you start a task or trigger an enabled workflow. It does not watch your desktop and act on its own.
The Free plan includes the agent workspace, OCR, and screen recording within free limits, plus limited writing assistant use. Pro raises those limits. Business is a custom team plan. When we checked, PixLab did not list Pro prices on the pages above, so check current pricing in the app.
How the controls work
For supported actions, Think-Act follows the policy you configure:
| Policy | What happens | PixLab's suggested use |
|---|---|---|
| ALLOW | The action proceeds | Scope and consequence are understood and no pause is needed |
| ASK | The run pauses for your decision | A starting point for write, send, publish, and delete actions while you test a workflow |
| BLOCK | The action is unavailable | The action does not belong in the task |
A few limits sit around that, and PixLab states them plainly:
- Policy applies to supported actions. PixLab says not every connector exposes the same permission controls.
- The activity you review after a run is, in PixLab's words, review evidence and not a complete audit trail.
- Checkpoints and recovery work only where the workflow supports them. PixLab says not every action can be undone, especially changes in outside systems.
- A Skill holds instructions only. It cannot add a connector, credentials, or permissions.
- The agent can act inside other applications only when a supported tool or connector gives it that ability and it is installed, enabled, trusted, and permitted. Without that, it prepares drafts and extracted data for you to use.
Local does not mean offline
Think-Act runs on your desktop, but PixLab's security page says context may be sent to a configured model endpoint or connector when a task uses one. The FAQ says processing location depends on the model, endpoint, connector, and workflow. It mentions private or local AI endpoints only "where your configuration supports" them, without listing which ones. If client data is involved, map the full data path before you treat Think-Act as an on-device tool. Our post on local vs remote MCP servers covers the same question for connectors.
Third-party MCP servers are your review job
PixLab says third-party MCP packages are not controlled or audited by PixLab, and that a local MCP server can run code on your device and reach files, credentials, networks, or application data depending on its permissions. Think-Act's ALLOW, ASK, and BLOCK rules do not replace the operating system's permissions or the package's own behavior. Before enabling one, check its source and maintainer, the command it runs, the tokens and folders it can use, and how it behaves on a small task. That is the same review we recommend for any agent host. See how to secure an MCP server for production.
What PixLab lists as roadmap
PixLab's FAQ lists cloud MCP connector registries and A2A remote-agent workflows as roadmap items, not current-release features. Organization-level policy and provisioning are also listed as planned. The downloads page says some organization-level controls remain roadmap items and tells teams to confirm what is available before planning a rollout. If you need central admin policy across many PCs today, ask PixLab directly.
One more thing we noticed: on October 10, 2026, the product pages showed placeholder images where app screenshots will go, and some pages say "Available for Windows and macOS" while the downloads page lists only a Windows installer. We went by the downloads page and the launch post.
OpenClaw and Hermes run a different kind of job
OpenClaw is a Gateway you run. It connects model providers, sessions, tools, messaging channels, the Control UI, and optional device nodes. It extends behavior with skills, plugins, and MCP servers. Sandboxing is off by default. When you turn it on, tool execution can move to Docker, Podman, SSH, OpenShell, or Crabbox while the Gateway stays on the host.
Hermes Agent is a process you run from a CLI, a messaging Gateway, an API server, or a supported editor. Tools come in toolsets you select, skills load on demand, and it can connect to local or remote MCP servers with per-server tool filtering. Dangerous commands go through an approvals setting (`smart`, `manual`, or `off`) with a hardline blocklist that stays on underneath.
All three use skills and MCP, so the overlap on paper is large. The difference is who is present and where the agent lives.
Side by side
| Question | PixLab Think-Act | OpenClaw or Hermes |
|---|---|---|
| Where does it run? | A Windows 10 or 11 desktop app | A host and sandboxes you configure |
| Who starts the work? | The person at the PC starts a task or enabled workflow | Chat messages, schedules, API calls, or a person |
| How are actions gated? | ALLOW, ASK, or BLOCK for supported actions | Tool policy, toolsets, deny rules, sandboxing, and approval settings in config |
| Skills and MCP | Local Skills and local MCP servers; cloud registries are roadmap | Skills plus local and remote MCP servers |
| Models | Tied to your PixLab account, plan, and usage limits | Provider choices you configure |
| Who maintains it? | PixLab ships and updates the app; you review connectors | You install, patch, and audit the runtime |
When you want both
They fit different parts of the same team. Someone in operations can use Think-Act to OCR a stack of scanned invoices, check the totals against a success goal, and record the process as a Skill. The job that has to run every morning, such as pulling new invoices from email and filing them in the accounting system, belongs in a self-hosted agent with a narrow MCP tool and an approval step before anything is posted. Our posts on human approval for AI agent actions and MCP permission retries and approval controls cover that side.
If you are comparing desktop agents, our posts on Muse Secure VM, Claude Code computer use, Simular Sai, Yellow.ai Nexus EDGE (enterprise EX Automation for IT/HR/Ops), and LUCI local screen context cover other ways vendors handle the same screen and approval questions.
Before you roll Think-Act out to a team
- 1Confirm the PCs run 64-bit Windows 10 or 11, and that Mac users can wait for the Q1 2027 release.
- 2Decide who holds PixLab accounts and which plan covers expected usage.
- 3Find out which model endpoint each workflow uses and whether client data may go there.
- 4Set write, send, publish, and delete actions to ASK while you test.
- 5List the MCP servers you will allow and review each one as third-party code.
- 6Keep credentials, private messages, and regulated data out of screen recordings unless your policy allows it. PixLab's FAQ gives the same advice.
- 7Confirm with PixLab which organization-level controls exist today before you write a deployment policy around them.
How this fits implementation work
OrchestriAI does not resell or operate Think-Act, OpenClaw, or Hermes for clients. MCP development builds narrow, reviewed tool servers that a desktop agent or a self-hosted agent can call. AI agent systems decide which jobs an agent may run and where a person approves. Systems integration connects those tools to the systems that hold the data.
Independence and limitations
OrchestriAI is an independent implementation provider. Product names are used for identification only. OrchestriAI is not affiliated with, endorsed by, or sponsored by PixLab or Symisc Systems, OpenClaw, or Nous Research / Hermes. This article summarizes PixLab's launch post (September 18, 2026) and its Think-Act Desktop Agent, integrations, security, downloads, and FAQ pages as checked on October 10, 2026, plus our existing OpenClaw and Hermes guides. We have not run Think-Act ourselves. Plans, platforms, and controls can change, so confirm current PixLab documentation before you install it on a machine that holds client data. This is not a security certification or a compliance conclusion.
