Construct Computer

Construct Computer is a general-purpose “AI employee” built around a persistent cloud workspace rather than a disposable chat session. Each agent gets a browser-based work desktop with Chat, Files, Browser, Terminal, Email, Calendar, Workflows, Memories, Activity and connected apps, plus an isolated sandbox terminal and a durable `/workspace` mount.¹ Files and selected state persist across sessions, while live browser runs and shell execution are bounded surfaces rather than an always-running personal VM.¹²

That persistent-workspace model is the core product differentiator. The agent can research the web, operate interactive sites, run code, create or modify files, send native email, act through connected SaaS tools, schedule future jobs and delegate bounded subtasks to temporary agents.³¹ Construct therefore sits between a conversational assistant, an automation platform and a cloud computer: the user assigns an outcome, while the agent decides which available execution surfaces are needed.

The integration surface is broad but mediated. Construct uses a live Composio catalog for supported SaaS applications and also supports custom MCP servers, direct Slack/Telegram/Discord access and private workspace apps.¹³ The current v2 platform is proprietary and does not yet have public registry submissions for apps/MCP packages, so extensibility is available but still more controlled than a fully open marketplace.¹

Construct’s persistence is more explicit than generic “memory.” The product stores long-term memories with source episodes, versioned assertions, temporal context and related entities; users can inspect, search, correct, forget or restore those memories.⁴ That makes memory governance materially better than an opaque profile the user cannot inspect. It does not eliminate memory error: the Terms warn that recalled context and summaries may be incomplete or inaccurate and that long sessions may undergo compaction.³

Recurring work has three layers. Calendar can schedule one-time or recurring agent jobs; Workflows encode reusable procedures; and the agent can turn repeated operational work into persistent files or private apps.¹⁵⁶ This is important because a fully agentic loop is expensive and statistically fragile when repeated. Construct’s own reliability documentation explicitly argues for resumable work, durable artifacts and shorter independent runs rather than trusting one extremely long chain to complete perfectly.⁵

Workflows are useful but not a full replacement for mature automation builders. Construct’s current workflow system is versioned and linear, and its own comparison documentation says branching, delays, approval steps, fan-out and subworkflows are not yet first-class workflow primitives.⁷ Conditional logic can be pushed into an agent step, but that trades explicit deterministic structure for model judgment. For high-volume stable automations, a traditional automation platform can still be easier to audit and cheaper to run.

Construct can also build private internal tools inside the workspace. The agent writes the app, validates it and publishes it into the cloud desktop while leaving source available for inspection and later modification.⁶ A strong reliability detail is the “last good build” behavior: draft source changes do not replace the current working build until validation succeeds. A failed update therefore does not automatically take the previously working app offline.⁶ This is not full application version history, but it is a sensible safety boundary for agent-authored internal tools.

The workstation is intentionally accessible to non-engineers. Unlike coding agents whose primary output is a branch or pull request, Construct exposes ordinary files, sent messages, app updates and Activity summaries that an operator or founder can inspect directly.¹ That makes the product more appropriate for cross-functional operational work than a repository-centric coding agent.

The current platform is still evolving quickly. The Terms call the service under active development and reserve the ability to modify, suspend or discontinue features.³ The product also exposes plan-specific caps on task steps, command runtime, parallel work and storage, so “persistent AI employee” should not be read as unlimited autonomous compute.

Construct now has clear public pricing. Lite costs $9 per month, with an annualized displayed price of $7.50 per month. It includes up to two AI agents, up to 50 steps per task, two parallel background jobs, 100 MB of cloud storage and a five-minute command runtime.⁸ A seven-day trial is available from the current product page.⁹

Starter costs $59 per month or a displayed $39 per month on annual billing. It provides six times Lite usage, an agent email address, background and scheduled tasks, a 30-minute advertised command runtime, 1 GB of cloud storage and up to five agents.⁸

Pro costs $299 per month or a displayed $199 per month on annual billing. It raises usage to 32 times Lite, allows deep runs of up to 1,000 steps per task, advertises one-hour command runtime, includes 3 GB of storage, enables Bring Your Own Key and allows up to 15 agents.⁸ BYOK currently supports provider credentials including OpenRouter, OpenAI, Anthropic, Amazon Bedrock and xAI; BYOK traffic is billed directly by the external provider and does not count against Construct’s bundled model budget.³⁴

Enterprise pricing is custom. The public pricing page adds workspace isolation, customer-controlled model keys, SSO, private deployment and custom MCP scope, and explicitly states for Enterprise that the workspace is not used as training data.⁸ That Enterprise-specific wording should not be generalized into a universal no-training promise for every lower tier, because the current Privacy Policy describes service improvement and model-behavior monitoring but does not state the same broad lower-tier commitment.⁴

The subscription model is fixed-price with internal model-cost budgets rather than per-message billing. The Terms make clear that each plan still has monthly and per-session limits covering budget, iterations, parallel jobs, schedules and storage.³ This is simpler for a small team than raw token billing, but “fixed price” does not mean uncapped usage.

There is one public documentation ambiguity around runtime limits. The pricing page advertises a 30-minute command runtime on Starter and one hour on Pro, while the current Terms and Privacy Policy say sandbox commands currently share a five-minute runtime cap.⁸³⁴ Those statements may refer to different execution layers, but the public documentation does not define the distinction clearly enough to assume that every shell command can actually run for 30 or 60 minutes. Users with long-running code or browser jobs should verify the live plan behavior.

The economic proposition depends heavily on the type of work. For ambiguous, cross-app work that would otherwise require a person to research, decide and update several systems, $59 or $299 can be inexpensive if the agent actually completes useful work. For deterministic high-volume workflows, conventional automation can remain cheaper and more reliable because it avoids repeated model reasoning.⁷

Construct has strong current first-party documentation but very little mature independent exact-product evidence. No substantive current exact-product review, controlled benchmark or mature verified-user corpus was located. This limits confidence in broad claims about completion rate, uptime, support quality and production reliability.

The launch itself generated real attention. Product Hunt ranked Construct Computer #1 on its August 23, 2026 launch day, and independent launch-history tracking recorded roughly 337 launch points and 37 comments that day.¹⁰¹¹ Launch traction is evidence of awareness and interest, not evidence that long-running agents reliably complete production work. Product Hunt commentary is also heavily maker-driven at this stage, so launch engagement should be treated as traction context rather than exact-product performance evidence.

A recent Japanese “5-minute review” describes the product positively as an AI employee that can run routine work, build internal tools and execute reusable workflows.¹² That review is useful as evidence that an independent writer inspected the shipped product proposition, but it is explicitly lightweight and does not provide a controlled reliability test or long-term usage study. Other current directory listings similarly confirm the product surface without adding a meaningful verified-user sample.¹³

Construct’s own reliability writing deserves credit because it is unusually candid about agent failure. The company explains why long multi-step tasks compound per-step error, recommends durable artifacts and resumable work, and explicitly notes that irreversible actions require careful supervision and idempotency.⁵ This is good engineering disclosure, but it is still first-party reasoning rather than measured Construct success rates across an external benchmark.

The most important safety limitation is external side effects. The Terms state that the agent may browse, execute commands, create files, send email, post on messaging platforms, schedule tasks, spawn temporary agents and call connected tools, and explicitly say Construct does not insert a mandatory approval before every external side effect.³ Users can inspect outputs, interrupt a running turn and answer agent questions, but the default model is supervision rather than universal pre-action confirmation. Construct’s own deployment guidance recommends running new jobs supervised before moving them to unattended schedules.⁷

Inbound access control is better separated. Slack and Telegram connections can be configured as open, approval-required, blocked or closed; guests are blocked by default and approval-required mode uses a sender review queue.¹⁴ This controls who can ask the agent to act, not whether each downstream action receives a separate human approval.

The platform publishes detailed data architecture. Workspace files are stored in scoped Cloudflare R2 prefixes; chat state uses per-agent Durable Objects; memory uses a user-owned MemorySpace; sandbox containers are scoped to the active workspace; Construct-stored BYOK and messaging credentials use AES-GCM; Composio manages its own OAuth credentials.⁴ This is useful technical transparency, although no current independent SOC/ISO or penetration-test evidence was located.

Third-party processing is broad by design. Connected-app prompts, parameters and results can pass through Composio, model prompts are sent to the provider serving the user’s plan, browser tasks can be processed by Composio Browser, and the platform itself uses Cloudflare infrastructure, Sentry, PostHog and Dodo Payments.⁴ BYOK changes the model-provider relationship but does not eliminate Construct’s need to retain conversation content for history and context.⁴

Data retention is relatively clear. Account and virtual-desktop data—including workspace files, memory, chat history, Activity history, Calendar and agent email—persist while the account is active and are deleted when the account is closed, subject to the policy’s stated operational/legal boundaries.⁴ Users can request access, correction, export or deletion and can inspect or delete individual memories from the product.¹⁴⁴

A separate privacy issue exists on the public marketing website. Construct’s Privacy Policy says the public site uses PostHog with unmasked session replay that can capture form inputs such as beta-signup email, on-page text, canvas, console logs and network request headers/bodies where captured, and that the site does not honor the browser Do Not Track signal for that analytics stack.⁴ This disclosure concerns the public website analytics surface, not a statement that private workspace content is broadly replayed by PostHog. It is still an unusually aggressive analytics configuration and privacy-sensitive visitors should know about it.

User-content ownership is comparatively straightforward. The Terms say users retain ownership of User Content and Construct receives a limited license to store/process/transmit/display it only as necessary to provide the service; Construct does not claim ownership of agent-generated content in the user environment.³

The platform is expressly provided “as is” and “as available,” and the Terms do not guarantee uninterrupted, error-free or secure operation or accurate agent output.³ That language is standard for a young SaaS product, but it matters more here because the product can autonomously change external systems. The current evidence therefore supports a detailed but low-confidence desk assessment rather than a public numeric score.

Construct is best suited to solo founders and small teams whose work is fragmented across email, CRM, project tools, documents, browser-only portals and ad hoc scripts. These users often have outcomes that are clear—research competitors, update a pipeline, prepare a report, follow up with leads—but cannot justify building a deterministic automation for every changing edge case.⁹¹

It is particularly compelling when work must continue after the user closes the browser. Persistent files, scheduled tasks, agent email and background jobs let a team delegate a recurring outcome rather than keep a chat session open.¹⁸ Users can then inspect the Activity history, files and messages afterward instead of relying only on an LLM’s narrative summary.

Small teams that want reusable automation without becoming automation specialists are another strong fit. Once a process stabilizes, it can be encoded into a workflow or turned into a private internal app rather than forcing the agent to rediscover the procedure every time.⁶⁵ This “reason first, then crystallize the stable parts” model is one of Construct’s better ideas.

Non-technical operators may also find the desktop metaphor easier to supervise than repository-centric agents. Files, sent email, calendar entries, app records and Activity summaries are understandable business artifacts; reviewing a pull request is not required for most jobs.¹

Construct is less suitable for deterministic, high-volume processes where every branch and failure path is already known. Current workflows are linear and lack first-class branching, delays, approval nodes, fan-out and subworkflows.⁷ Zapier, Make, n8n or custom code can be more predictable for that class of work.

It is also a weak fit for risk-sensitive organizations that require a mandatory approval before every consequential external action. Construct provides interruption, activity visibility and inbound sender controls, but the current Terms explicitly say there is no universal pre-side-effect approval gate.³ A new workflow that can send messages, change CRM data or submit forms should be supervised until its failure modes are understood.

Organizations needing audited enterprise assurance should perform additional diligence. Enterprise offers SSO/private deployment/custom MCP scoping and stronger data-positioning language, but no mature third-party security or enterprise-review corpus was located.⁸

Finally, teams with very long shell workloads should verify execution limits before purchase because the current pricing page and legal/privacy documentation describe runtime caps differently.⁸³⁴