Circe · Product research · 19 September 2026
A real need. A crowded set of solutions. An unproven opportunity in helping people delegate work and step back into their own computers.
Assessment based on current primary documentation and research, plus the Circe discussion. Competitor capabilities below are documented claims, not a hands-on comparison. Images are generated concepts. No Circe implementation was changed.
Worth a focused experiment. Too early for a full product pivot. The original moment contains useful evidence: you wanted something done on a laptop without abandoning what you were doing on your desktop, then wanted to inspect and modify the result.
That establishes one person's need. It does not establish frequency, a market, willingness to pay, or an advantage over existing products.
“I can work through any of my computers, delegate parts of that work, and take over without losing my place.”
Circe would earn its place by making that exchange dependable. Voice, device routing, automation and remote viewing would support it. Each is also available elsewhere.
The earlier architecture briefing describes a possible implementation. It should not be read as proof that the product direction is correct. This assessment puts demand, alternatives and interaction quality ahead of a wider build.
Who already serves parts of the need?
The presence of many startups shows there is substantial technical and commercial activity. It does not tell us which product Circe should build. Browser infrastructure can succeed selling capacity to developers even when a general personal assistant remains unreliable.
| Category / example | Documented capability | Implication for Circe |
|---|---|---|
| Remote access RustDesk | Open-source remote control with self-hosting options. | The baseline is already strong: connect and do the task yourself. Circe must beat its total effort for the chosen work. |
| Personal assistant Claude Dispatch + computer use | Assign tasks from a phone to a desktop. Computer use can operate apps and the browser. The current announcement lists macOS and Windows and requires the desktop awake and running. | Remote delegation is already a direct product offer. “Talk to an AI that uses your computer” is not an empty category. |
| Personal agent runtime OpenClaw | A gateway runs the agent and calls paired nodes when their tools are needed. Remote devices can expose commands and local browser capabilities. | Multiple owned machines and extensibility are already being addressed. Better everyday interaction must be demonstrated. |
| Browser execution Browser Use | Supports using a system Chrome profile for authenticated tasks. Its documented approach can require closing Chrome first. | “Uses my logins” and “coexists with the exact browser session I am using” are separate promises. Evaluate both. |
| Hosted browser infrastructure Browserbase / Stagehand | A published template supports pausing for human input, resuming an agent and watching through Live View. | Human handoff alone is not unique. Hosted browser work is also different from operating the user's actual laptop. |
| Desktop execution infrastructure Cua | Driver operates existing machines; Sandbox supplies isolated computers. Code, structured tools and GUI actions can be combined. | Assess reusable execution components before owning every OS mechanism. An SDK remains one dependency, not the whole user experience. |
| Desktop automation product Simular | Markets a developer-oriented computer-use platform and repeatable automation built on its Simulang execution language. | Broad app coverage and repeatable workflows are another existing product direction. Vendor reliability claims require testing. |
| Business process automation UiPath Maestro | Coordinates automation, AI agents and human interactions across business processes. | Enterprise process ownership is a different buyer and sales motion. General computer control does not automatically give Circe that business. |
Representative comparison, not an exhaustive market inventory or a feature ranking. Documentation reviewed on 19 September 2026. Product plans, OS support and preview status change.
An online node may have no unlocked graphical session. A command may reach the machine but miss the desktop, account, browser profile or open document the user meant. Pairing, OS permissions and sleep behavior are part of the product.
Cua documents Windows interactive-session boundaries and different background action mechanisms on macOS, Windows and Linux. Background operation is explicitly best effort, with some actions requiring foreground access. Source
A click can hit the right button while the request fails, the wrong account is selected, or a save remains pending. Completion needs observable evidence appropriate to the task: the reopened document, the saved value, or the resulting artifact.
This is an engineering requirement proposed for Circe. A fluent completion message cannot substitute for checking the result.
You can type, move windows, change a document or walk back to the laptop while the agent acts. “Take control” must pause conflicting automation and discard stale assumptions. “Continue” must inspect your changes.
For an initial version, serialize desktop input on each machine. Add background concurrency only where its behavior is demonstrated.
After a connection drops, did the remote app receive the final action? Retrying a submit or send can duplicate it. UI actions often lack a reliable request identifier or an undo operation.
Record uncertainty, inspect the target state, and ask when inspection cannot resolve it. Do not display “Stopped” until the remote worker has acknowledged that it stopped; disconnection is a different state.
Web content, documents and tool descriptions can contain instructions the user never authorized. Scope execution to the user's request and approved resources, including the account and destination.
WebMCP's security guidance discusses origin isolation and sensitive tool handling. Structured tools improve access but do not establish intent or permission. Source
OSWorld 2.0 contains 108 extended workflows. Its published evaluation reports a best full-completion score of 20.6% under its 500-step configuration. The tasks have a median human duration around 1.6 hours.
That is not a predicted failure rate for Circe's short tasks. It is evidence that long tasks require separate evaluation; short demonstrations cannot establish broad autonomy. Benchmark and methodology
WebMCP lets participating websites expose structured actions through JavaScript and annotated forms. Chrome documents it as a proposed web standard in origin trial. It can reduce the need to infer a site's controls, but requires site participation and an appropriate browser context. Chrome documentation
For Circe, treat it as one execution option alongside service APIs, browser automation, OS accessibility and visual control. Select by the required account, actual app state and action quality. An API that acts on a different account or stale copy is not automatically the better path.
It does not solve native desktop apps, machine sleep, conflicting human input, task recovery or verification. Building the product around universal WebMCP availability would be premature.
Visual exploration
The images use one illustrative design task so the interface differences are easy to compare. Circe is not being defined as a portfolio builder. These are independent image concepts, not working interfaces or screenshots of shipped capabilities.
My hypothesis is that the small summon panel should expand into the shared session when needed. The result view is useful for work with a natural artifact. These can eventually coexist, but the first experiment should test one complete interaction rather than build all three shells.
Always show the machine and whether the view is live or stale. Give voice and typing the same capabilities. A machine may be offline, waiting for access, working, waiting for you, or finished with evidence. Use ordinary language for those states. Avoid a cheerful generic spinner.
My candidate initial audience is people who already do meaningful work across two computers. That includes some developers, creators and technical solo operators. This is an audience hypothesis, not evidence that all those groups have the same needs.
Start with work that is tied to an existing machine: open local projects, configured tools, signed-in accounts and long-running processes. If a fresh cloud browser completes almost every requested task, the personal-machine thesis gets weaker. If users mostly want direct control, ship a useful remote-work product and keep autonomy modest.
Being open source and provider-neutral may help users choose Circe. Neither is sufficient protection from competition. A defensible advantage would have to emerge from superior setup, reliable task coverage and frequent use in a specific audience. We have not measured those yet.
These inputs are illustrative, not Circe measurements. A slow agent may still help if it frees attention. A fast agent may waste attention if it needs constant checking.
min
The task returns time to the user under these assumptions.
Manual active time minus explanation, supervision, verification and recovery. Measure elapsed completion time, monetary cost and interruption burden separately. Setup and permission friction must also be amortized over real use.
The next step
Run a proposed two-week pilot with roughly six to eight people who already use two computers. Begin with the actual operating systems and applications they use. This sample can reveal friction and repeated needs; it cannot establish market size.
Several participants independently return for repeated work; delegation saves active time against their baseline; corrections are cheap; target and completion states remain understandable.
Choose numeric success thresholds after observing baseline task costs and before the comparison, rather than inventing impressive universal percentages.
Requests happen rarely; most work could run in a generic cloud session; people routinely abandon the agent for direct control; setup dominates use; or one app/workflow accounts for nearly all value.
A concentrated workflow may justify a narrower product. Technical breadth is not a reason to ignore that result.
| Own in Circe | Evaluate for reuse | Postpone |
|---|---|---|
| Named machine and account scope; task lifecycle; understandable status; safe exchange of control; result verification; resuming after user changes. | Browser execution engines, desktop drivers, screen transport and networking. Compare support, licensing, integration effort and failure behavior before selection. | A universal agent model, every OS at once, phone-wide control, a new browser, generic business workflow authoring and “anything on any device” claims. |
The browser and desktop work already underway is useful as an execution foundation. Its next milestone should be a coherent user task that survives interruption and correction. Tool count alone cannot establish readiness.
Proposed responsibility chain: user request → resolved machine and scope → durable task → appropriate execution adapter → observed result → review or direct takeover. Keep voice transport separate from task lifetime, and keep execution on the selected node across reconnects. This is a direction for an experiment, not a finalized architecture.
We have not compared these products hands-on, measured Circe's task success, established setup cost across OS pairs, interviewed users, or tested willingness to pay. The research makes the alternatives and constraints clearer. It cannot honestly answer those questions for us.
Recommended next action: build a corpus of real requests from your own two-machine use, then compare one complete delegate–takeover–resume flow against doing the work directly.
Generated UI images are illustrative and contain invented content. They were made with the built-in Image Gen tool, grounded in Circe's crisp panel and control geometry. No computer-control or browser acceptance tests were run for this briefing.