Genspark AI Workspace

Genspark's agents write into systems it does not own. Its Google Workspace add-on, listed on the Google Workspace Marketplace by Mainfunc and updated 26 May 2026, works inside Google Docs, Sheets and Slides. The listing says Genspark "reads what you're working on (and any text, cells, or slide you have selected) and turns plain-language requests into real edits, formulas, analyses, slides, and on-brand visuals, right inside the file you already have open."

Named write actions follow: inserting generated images into a document, writing VLOOKUP, XLOOKUP, ARRAYFORMULA, QUERY and Apps Script formulas into cells, creating native Sheets charts placed without overlapping existing ones, applying conditional formatting and number formats, deduplicating and reshaping ranges, and generating summary tabs. Writing a working formula into a customer's spreadsheet is a state change in a system Genspark does not own.

The Apple App Store entry for Genspark AI Workspace, also vendor published, names Super Agent, AI Slides, AI Sheets, AI Docs, AI Developer with "full GitHub integration for development workflows with code editing and pull requests", and AI Meeting Notes. Opening a pull request is another write into a third party system. No full integrations directory is listed.

Sourceworkspace.google.com/marketplace/app/genspark/592480783716read 2026-09-12

Super Agent runs multi step, multi agent work. The agent "breaks your request into steps, chooses the right tools, and carries the work through to a finished result", and "coordinates the right agents and tools automatically." Parallel execution with dependencies is named: "Work in parallel. Run several tasks at once in a single project, and let each one build on the others' results." That is fan out with result dependency between branches rather than a chained prompt.

Durability is stated: "Sharper on long-running tasks. It reasons through every step without losing the thread. Start it, close your laptop, and come back to finished work." A run that survives the client disconnecting is a server side runtime. The comparison table draws the same line against the prior generation, which handled multi step tasks "primarily in sequence". No BPMN or named runtime is published.

Sourcegenspark.ai/helpcenter/super-agentread 2026-09-12

SecondBrain is described as "a personal knowledge and memory system, a unified hub that learns from your emails, meetings, files, and connected apps." It is a structure over the customer's corpus, maintained rather than assembled per run: "once connected, data begins syncing automatically in the background", and "SecondBrain begins building your personal context layer automatically."

Named sources are Gmail, Google Calendar, Outlook, Slack, Notion, HubSpot, Salesforce, Google Workspace and Microsoft 365, plus manual upload up to 500 MB and auto saved past Genspark activity. "Some integrations sync on a schedule rather than instantly" confirms a standing sync rather than a fetch at question time.

It stays queryable and scales past the context window. The published play "searches across all your connected sources to find any file, email, or note", and Super Agent "retrieves SecondBrain data" at run time. The result is a retrieval layer that persists between sessions and is searched on demand.

No embedding, graph or index technology is named; the behavior is documented and the mechanism is not.

Sourcegenspark.ai/helpcenter/secondbrainread 2026-09-12

Access control governs what Claw may reach, but no review and approve gate sits over what it does. The design posture is stated: "Claw is designed with human oversight in mind. External actions like sending emails or posting messages require permission."

Direct message access defaults to pairing mode, where "only contacts you've explicitly approved can message Claw", with a published warning that an open policy lets "anyone who knows Claw's address send it messages and prompt it to act on your behalf", plus an Allowed Senders list on the Claw email address. Each service is authorized through an explicit login flow, under a published minimum access principle.

Permission is granted at connect time, not per action. Once a service is connected the session persists and Claw "reads and sends emails, manages calendar events, reviews PRs without you opening the app", and scheduled tasks and Heartbeat run unattended by design. There is no approval queue, confirmation step, or policy layer deciding which actions need sign off, and nowhere for an administrator to configure one.

The vendor's own guidance shows no gate exists. For actions that are hard to reverse ("sending bulk emails, posting to social media, deleting files, or executing financial transactions") it advises "ask Claw to show you its plan before it starts", and to review planned actions where prompt injection risk is higher. The FAQ itself raises the residual risk ("What if Claw does something wrong while I'm not watching?"), and advice to ask the agent to pause is not a mechanism.

Sourcegenspark.ai/helpcenter/genspark-claw, Channels, Security and FAQ sectionsread 2026-09-12

The Security & Compliance section of the Business page states "SOC 2 Type II (Certified)" and "ISO 27001 (Certified)" as completed, marks ISO 42001 and GDPR as "In progress", and links a trust center at app.vanta.com/mainfunc.ai/trust. The vendor separates certified from in progress in its own copy, and the two certified entries are unhedged.

The SecondBrain page repeats it independently: "Genspark maintains an enterprise-grade security framework and is certified with SOC 2 Type II and ISO 27001." On access, the Team Plan documents "member roles, usage analytics, SSO/SAML, invoices, and connector management" under "centralized admin & billing controls", with an SSO FAQ entry: identity integration, a named role model and administrator control over which connectors a team may authorize.

One first party contradiction stands. Genspark's Google Workspace listing says the add-on can access "never your Drive, Gmail, or other files", while the permission manifest on the same page requests a Gmail label scope, a Drive file scope and personal info. Both cannot be true. That is a permission scope broader than the function requires, a question about scope discipline rather than a contradicted attestation.

Sourcegenspark.ai/business Security & Compliance section, corroborated at genspark.ai/helpcenter/secondbrainread 2026-09-12

Run visibility and failure diagnosis are documented for Claw, but there is no audit trail of what the agent did, when, and on whose authority. The Schedules tab shows per task run state ("each task shows its name, schedule, and enabled/disabled status"), and "if a task fails, a Last Run Failed button appears", with Claw diagnosing the error and suggesting a fix. A Diagnose tool "inspects system health, checks running services" and can restart components, and Terminal and Files give direct inspection of the machine. On the Team Plan, usage analytics provide administrative reporting.

Reporting is not auditing. These artifacts answer whether it is running and whether it broke. Nothing records which action Claw took in which connected service at what time under which authorization, nothing is exportable, no retention period is stated, and there is no administrator view across a team's agents; since "Claw is never shared", no cross user oversight surface exists by design.

The vendor names the gap itself: "there's no built-in way to share or export your Claw chat history", since "the conversation lives on your own cloud computer and can't be linked or shared directly." The suggested workaround is to ask the agent to write a summary, which is the agent reporting on itself.

The Team Plan MSA gives Genspark the right to audit customers' records relating to their use of the services annually; that is the vendor auditing the buyer, not a record of what the agent did.

Sourcegenspark.ai/helpcenter/genspark-claw, Schedules, Settings and Sharing sectionsread 2026-09-12

Claw has a real persistent memory layer: "Claw has a persistent memory system that works across sessions and channels." Say "remember this" and Claw "writes it to long-term memory and applies it automatically in future conversations." Long-term memory is shared across all channels while individual threads stay separate per channel, and memory is what survives automatic context compression on long threads.

Scope and lifetime are documented. Each user gets their own isolated cloud computer and "Claw is never shared, each Claw belongs to one person"; a Team plan buys seats rather than a shared instance. Memory lives for the life of the cloud computer: Delete VM "permanently erases conversation history, memory, files, and credentials", and on cancellation the machine is recycled and all data on it permanently deleted after the billing cycle.

What is stored is preferences, rules and context the user dictates on command, not a record the agent writes as a consequence of its own runs. And while the write surface is explicit, no memory inspection surface is documented: no list, editor or per item expiry, and no way to see or delete a single memory short of destroying the whole machine. That is documented cross session persistence without governance over the store. SecondBrain, the separate knowledge hub, is not part of this memory.

Sourcegenspark.ai/helpcenter/genspark-claw, Memory section and FAQread 2026-09-12

The agent can run on the buyer's own hardware. Genspark Claw runs "on a cloud computer (always on, 24/7) or locally via the desktop app (no subscription required, uses your own machine)", downloadable for macOS or Windows, and the vendor notes it therefore "has access to your file system".

That makes it a named, self serve option to run the agent in the customer's own environment. The cloud mode is single tenant: it "provisions a dedicated machine just for you" with its own CPU, memory, storage and a fixed IP address, and "each user gets their own isolated cloud computer, nothing is shared between users."

The option covers one agent, not the platform. Local mode runs Claw; the rest of the workspace, Super Agent, AI Slides, AI Sheets, SecondBrain and AI Drive, remains cloud only on Microsoft Azure in the United States. The vendor lists what local mode gets (chat, channels, skills, memory) as a subset that "will continue to gain features over time."

There is no region list, EU option, region picker, VPC or on premises server deployment, and the privacy policy's transfer clause to the United States still stands. A European buyer with a localization obligation is served only if Claw on a laptop is all they need.

Sourcegenspark.ai/helpcenter/genspark-claw, Quick Start Option B and Plans sections, against genspark.ai/privacyread 2026-09-12

Customers get prebuilt work in three forms. First, a named set of prebuilt agents the customer chooses among: the Help Center indexes Super Agent, AI Slides, AI Sheets, AI Docs, Genspark Code, AI Meeting Notes, AI Pods, Call For Me, Custom Agent, GenMail, Realtime Voice, Genspark Design and Genspark Claw as separate product articles. Each is a whole product: delete AI Slides and AI Sheets still does its own job, and deleting Call For Me leaves AI Meeting Notes unaffected. These are not tiers of one agent or modules the customer wires together.

Second, a browsable catalog. The Tools index at genspark.ai/tools carries roughly 150 individually addressable tool pages under Writing & Content, Design & Visuals, AI Models, Audio & Video, and Business & Productivity, each its own adoptable unit, such as AI CRM Builder, AI Pitch Deck Generator, AI Phone Call and AI Invoice Generator. Under Toolkits, the Business page counts "150+ reliable in-house toolsets designed specifically for complex business operations."

Third, and strongest because the customer authors into it, Skills: "Teach Super Agent a workflow once, and it's saved as a skill. Run it anytime, and share it with your whole team", with the vendor pointing to an official skills library alongside customer authored ones.

Sourcegenspark.ai/helpcenter and genspark.ai/businessread 2026-09-12

Work reaches Genspark Claw without a person invoking it: on a schedule, on a monitoring interval, or by arriving in a queue. The Schedules tab "creates recurring tasks that Claw runs automatically", with "no manual triggering required", configured as a plain time or "as a cron expression for advanced scheduling" with a timezone, and "Claw will run it at the specified time, even if you're not online." Each task carries a name, schedule and enabled or disabled state.

Heartbeat covers events, and it is the stronger of the two. Toggled on, Claw "proactively monitors your inbox, calendar, and other connected services at a regular interval and acts on anything that needs attention", even without a message. The vendor draws the distinction itself: Schedules is for tasks with a clear definition, Heartbeat for ongoing awareness without a task defined upfront.

Inbound queues are documented too. Claw holds a dedicated, always connected email address at the genspark.email domain with an Allowed Senders list, so mail addressed to the agent is an inbound work queue, and inbound messaging runs across WhatsApp, Telegram, LINE, Slack, Microsoft Teams, Discord, Signal, Google Chat and Feishu, with conversations kept separate per channel. SecondBrain's background sync is an ingestion schedule and starts no run.

Sourcegenspark.ai/helpcenter/genspark-claw, Schedules and Channels sectionsread 2026-09-12

Genspark lets the customer choose models, on top of automatic routing. Super Agent routes subtasks automatically across providers including OpenAI, Anthropic, Google, xAI and ElevenLabs when the user does not choose, and the user can choose. The vendor's own credit guidance only makes sense if selection is the user's: "choose a lower model tier when you don't need maximum power."

The selectable set is enumerated per agent.

Team Plan chat names Claude Opus 5, GPT-5.6, Gemini 3.8 Flash, Grok 4.6, DeepSeek V4 Pro, Kimi K3, MiniMax M3 and GLM-5.3; image names Recraft V4.1, Ideogram V4, Qwen Image 3, Seedream 5.0 Pro, GPT Image 2, Nano Banana Pro, Nano Banana 2, Seedream v5 Lite and Flux 2 Pro; video names Gemini Veo 3.1, Seedance v2, Kling V3, MiniMax H3 and PixVerse V6; audio names ElevenLabs, MiniMax, Mureka, Google, Microsoft and CassetteAI.

The headline is "70+ AI models including ChatGPT, Claude, and Gemini." The Tools index adds a third selection surface, per model landing pages under an AI Models heading such as genspark.ai/tools/claude-sonnet-5, /grok-4-5, /gpt-image-2, /seedream-5 and /nano-banana, each an addressable entry point to a named model.

Azure, AWS and GCP on the subprocessor list are infrastructure, not model providers. No admin level model policy for a team is published.

Sourcegenspark.ai/business Team Plan section and genspark.ai/helpcenter/super-agentread 2026-09-12

There is no published interface for calling the platform. No REST or GraphQL API, SDK in any language, developer portal, MCP server, A2A agent card, endpoint or authentication scheme is documented. The Help Center index, 28 product articles and 6 account articles, has no developer, API or integration build section, and the Business page footer has no developer entry.

Every documented surface points outward. The Chrome Extension, Microsoft Office Plugin, Salesforce Integration and the SecondBrain connector set (Gmail, Calendar, Outlook, Slack, Notion, HubSpot, Google Workspace, Microsoft 365) all let Genspark reach into other systems. A bridge exposing the customer's systems to agents does not expose the vendor's platform to outside callers.

AgentBase, in a product family with Custom Agent, Skills and Hub, most likely builds agents inside Genspark rather than opening it to outside callers. GenTerminal is a separate product at genterminal.ai, a terminal rather than an interface for driving this platform.

Sourcegenspark.ai/helpcenter and genspark.ai/businessread 2026-09-12

Genspark publishes no way for a customer to evaluate a change: no evaluation harness, scored test cases, quality gate in a release path or controlled post deployment experiment. Nothing in Super Agent, SecondBrain, the Business plan or the Help Center offers a customer a place to test an agent before or after relying on it.

Three things look like evaluation and are not. The new Super Agent "checks own output and build on earlier progress", which is the agent checking itself inside a run, with no result the customer reads and no comparison against anything.

Genspark's engineering blog describes accumulating "real-world examples of successful tool selection and model routing" internalized "through reinforcement learning and prompt playbooks", a genuine optimization loop, but over the vendor's engine across all users rather than this customer's agent, and the customer neither invokes it nor receives a result.

And the guidance to "nail the prompt in AI Chat first", shaping and refining a request in AI Chat before handing it to Super Agent, is advice about drafting a request, not a harness.

No bias or fairness audit is published, and ISO 42001 is listed as in progress on the Business page, which is not a shipped surface.

Sourcegenspark.ai/helpcenter/super-agent and genspark.ai/businessread 2026-09-12

Browser use, remote computer control and local computer control are all documented for Genspark Claw. Browser: connecting a service "opens a remote desktop window", which is "the browser on your cloud computer, already navigated to that service's login page", and the session persists so Claw can act in that service afterwards.

Remote computer control: a Remote Desktop button "opens a browser-based view of your cloud computer directly", with a resettable password, a Terminal tab giving "direct command-line access to your cloud computer in the browser", and a Files tab that is "a browser-based file manager".

Local computer control: the desktop app "runs Claw directly on your own computer" for macOS or Windows, and the vendor's security note says it "has access to your file system", with a workspace folder offered as "soft guidance, not a hard boundary".

The buyer directs and observes it. The buyer chooses a tier (Lite, Standard, Powerful), provisions a dedicated machine with its own CPU, memory, storage and fixed IP, watches it through Remote Desktop, drives it through Terminal, and manages it through Refresh Status, Update Configuration, Diagnose, Force Restart and Delete VM.

The vendor acknowledges the fragility, a strong sign of real interface operation: it warns that when Claw "browses websites, reads emails, or processes documents" malicious content may attempt prompt injection, and advises reviewing planned actions in less controlled environments. Terminal and shell access is autonomous execution rather than interface operation, and Call For Me is a telephone keypad path, not computer use.

Sourcegenspark.ai/helpcenter/genspark-claw, Services, Home, Terminal and Security sectionsread 2026-09-12