Personal Agents Platforms

Related: For agents shared through Slack or Discord, team agent platforms covers shared workplace agents; agent frameworks covers the components used to build them.

Executive Summary

A personal agent needs more than a chat window. It needs access to the right files or accounts, a way to take actions, useful context between tasks, and a clear answer to what happens while its owner is away. This report compares 45 platforms and deployable projects, checked September 16, 2026, with Instinct added September 23. It separates hosted services, device assistants, self-operated gateways, experimental deployment stacks, and specialized hardware agents rather than ranking them by repository stars.

The consequential choice is where the work runs. ChatGPT Work offers distinct local and cloud tasks; Claude Cowork now defaults to a cloud environment while device work still needs a connected desktop. A local OpenClaw gateway can call remote models and services, so self-hosting alone does not keep all data on the machine.^[1][2][3]

Six substantive entrants join this edition: ChatGPT Work, Cowork, Meta Muse, Agent Zero, Letta, and Grok Bot. Waitlists, deprecated software, deterministic assistants, and developer-only components remain visible in the context section, but do not inflate the member count. Inclusion of experimental code means that it is available to evaluate, not that its reliability or security has been certified.

---

Market Definition

For this comparison, an included product or project must provide a usable agent experience or deployable implementation that accepts a person's goals, takes actions through tools or a computer, and supports continuing context, repeated work, or a persistent working environment. A specialist scheduling or device agent can qualify without being a universal assistant. Team features alone do not disqualify a personal assistant. Developer harnesses and systems organized primarily around repository configuration and reviewed team changes remain adjacent context rather than automatic members.

The categories describe who operates the runtime and what the product actually does:

• Hosted and cross-device services: The provider delivers an assistant application, with hosted execution or connected local environments where offered. The user authorizes accounts, supplies context, and reviews output.
• Device assistants and installers: A desktop or phone is an important part of the experience. Local execution and local inference are different properties.
• Self-operated gateways and workbenches: The user installs software, chooses models, configures tools, and owns upgrades and recovery.
• Operator reference stacks: Available deployment code with explicit alpha or experimental status.
• Specialized devices: An agent acts through an Android phone or microcontroller, often using a model endpoint elsewhere.

The last substantive edition was June 11. Every retained matrix row was checked against current primary material, including GitHub documentation and current vendor guides. Tool access through MCP is evaluated as an integration path, not a security guarantee. Prices below are published US-dollar monthly terms where verified, not normalized cost-per-task estimates. Trials, subscription allowances, hosting, third-party services, and bring-your-own-key inference are distinct charges. Capabilities are documentation-verified; no common benchmark was run across these products.

---

Comparison Matrix

Hosted and cross-device services: 15 members

Device assistants and installers: 5 members

Self-operated gateways and workbenches: 20 members

For these rows, evaluate software licensing separately from model, server, storage, and integration bills. A release binary is not a bundled inference service.

Operator reference stacks: 2 members

Specialized devices: 3 members

---

How the Main Workflows Differ

A computer that produces artifacts

Manus, ChatGPT Work, Cowork, HappyCapy, and Agent Zero are especially relevant when the output is a report, workbook, presentation, repository, or edited document. They are not interchangeable computers. Manus distinguishes its default temporary task sandbox from an optional persistent Ubuntu Cloud Computer, whose files and processes survive individual tasks. Cancelling that computer can delete its files even though generated outputs already saved to a chat remain.^[63][6]

Work cloud tasks use hosted infrastructure, while local Work accesses approved device resources. Cowork's current cloud default can run schedules while a laptop is off, but local files, browser, and computer use need an open, connected desktop and the appropriate session. Both products require attention to which surface actually owns the task.^[64][2]

Agent Zero is the operator-controlled alternative in this group: it combines a Docker Linux desktop with LibreOffice, browser annotation, and editable project context. Its host connector is optional; terminal execution can reach locations available to the Launcher user beyond the selected folder. This makes it useful for real computer work while changing the relevant trust boundary.^[65][66][67]

HappyCapy packages hosted compute, skills, automations, and email into plans. The distinction to verify is which quotas and features the chosen tier actually includes. Its pricing page still labels GUI agent teams a research preview and iOS access as early access, rather than treating every advertised surface as generally available.^[16]

Meta launched Muse on September 8 with US rollout on mobile and web, plus WhatsApp access. Its launch describes a dedicated Secure VM and a separate Sentinel that reviews outbound activity. These are vendor-described controls, not an independent security assessment. Confidential VM, intended to prevent Meta from accessing its contents, is a later-year plan; training opt-out is a separate current setting, not a default no-training promise.^[13]

Grok Bot: persistent roles on one personal computer

Grok Bot qualifies as a hosted personal agent because it combines continuing context with real tool actions and a persistent working environment. Cloud tasks continue while the laptop is closed. Bots can coordinate, retain separate conversations, and reuse learned workflows.^[11] Skills capture procedures; routines supply schedules or supported event triggers. Demonstration-based teaching is gradually enabled, so it should not be assumed available in every account.^[68]

The important boundary is one computer per user, shared by all that user's Bots. Files and signed-in browser sessions are shared; separate Bot screens are work surfaces, not credential isolation.^[69] Self-serve Teams access also differs from Enterprise administration: destination controls and enforced team Auto Review are Enterprise features.^[70]

This makes Grok Bot worth shortlisting for recurring personal research and cross-application preparation. Test whether one Bot's accounts are appropriate for every role on the same computer. For shared operational review, compare the team workflow described under Tembo Agent Studio below; the products organize responsibility differently.

An assistant where messages already arrive

Poke, Instinct, Lindy, Ako, and Majordomo reduce the distance between a request and an agent. Poke uses familiar messaging clients; Instinct is invite-only text or call, with testers reporting real-world bookings and also unauthorized email and inbox retention after disconnect; Lindy now spans Slack, iMessage, email, meetings, and browser work; Ako supplies an assistant identity across WhatsApp, Slack, and email; Majordomo uses an iOS interface to a hosted agent computer. These interfaces still authorize cloud services to act on connected accounts.^{[22][23][71][7][17][26]}

Poke also has a changed owner: Cognition announced its acquisition of Interaction, Poke's maker, on July 23, 2026 and said the service would continue. That is a current ownership fact, not evidence that future integration plans have already shipped.^[72]

Blockit deliberately narrows the job to scheduling. Its site describes policies ranging from review-first to automatic scheduling and rules based on the person or meeting. A narrow agent can be the better fit when the desired outcome is reliably arranging meetings rather than providing a general computer.^[25]

Workspace subscriptions and preview gates

Gemini Spark's help page is more specific than a broad launch headline: it calls the service experimental, requires an eligible personal account with Google AI Pro or Ultra, and lists age, activity-setting, and regional restrictions. The overview and help text differ on business availability; this report follows the help page's explicit current account restrictions. Its remote computer can continue without the user's device, while local Chrome work needs that device available.^[73][9]

Scout remains a Microsoft Frontier preview requiring enrollment and the documented licensing and app-policy prerequisites. It combines local files and tools with Microsoft 365 Work IQ and can use memory and periodic follow-ups. A Microsoft 365 account alone is not sufficient evidence that the user can deploy it.^[27][28]

Perplexity Personal Computer now supports Windows as well as Mac. The Mac setup guide requires macOS 15 or later, with Comet for browser work and phone-based approvals; a Mac mini is recommended rather than compulsory. Accessibility, screen/audio recording, and Full Disk Access are substantial permissions. Local work requires the host to remain available.^[30][74]

Its newer compute modes deserve a separate hardware check. September 1's Hybrid Compute combines cloud planning with local-model processing and a privacy gate that can mask, block, or request approval for outgoing information. It requires Apple Silicon, macOS 15+, and at least 24GB unified memory, with Pro, Max, or Enterprise access. September 14's Portable Computer for Windows runs the model, agent harness, orchestrator, and scheduler locally on supported NVIDIA RTX hardware with at least 24GB VRAM. The vendor says local work consumes no Computer credits and cloud transfers require permission. Connected services and approved cloud calls still cross the device boundary; neither mode proves that every workflow stays local.^[31][32]

Persistent memory and a self-operated gateway

OpenClaw remains a broad personal gateway with messaging, plugins, and configurable tools. Its README explicitly describes the data path: prompts can go to model providers, and conversations pass through the selected messaging service. Its default host execution must be considered separately from optional sandboxing.^[3]

Hermes pairs continuing memory with learned skills and a messaging gateway. Its supported terminal backends include local execution, containers, SSH, and hosted environments, so a recommendation must specify the backend. NanoBot offers a smaller Python-oriented deployment with CLI, TUI, Web UI, tools, and a background gateway; a running process is what makes its scheduled work continue after a client disconnects.^[38][39]

Letta is a current alternative for persistent agents with editable memory, skills, remote environments, and messaging channels. It is also the relevant migration context for the archived LettaBot. Lobu now combines an event-sourced context layer with persistent specialists and durable automations; evaluating it only as an old Slack bot misses the connector permissions and context provenance that now structure its product.^[21][75][54]

Regional channels and extension models

AstrBot and QwenPaw warrant evaluation when Chinese messaging systems are central. AstrBot's README lists QQ, WeChat-related surfaces, Feishu, DingTalk, and international channels, while still marking WhatsApp as forthcoming. QwenPaw combines messaging with AgentScope-based tools, skills, memory, and local or hosted inference. Verify the precise adapter and account setup required for the intended channel rather than assuming every logo has equivalent access.^[40][45]

NanoClaw takes a different extension approach: channels and model-provider variants are selected through documented skills/branches. The current version runs agents in Docker containers. Choosing a repository fork or provider variant is an operational choice that affects upgrades and compatibility, not just a checkbox in a universal feature matrix.^[42]

MicroClaw now has both a server runtime and a native Work application. Its stable branch is the recommended production path; the main branch can break compatibility. Windows/Linux desktop previews should not be conflated with supported Apple Silicon Work builds or the broader server platforms.^[51]

Small runtimes, specialists, and experiments

ZeroClaw, NullClaw, ZeptoClaw, Moltis, and IronClaw offer different combinations of compiled runtimes, channels, memory, and tool isolation. Language choice alone does not establish lower end-to-end cost or stronger security. For example, IronClaw's WASM tools, Docker jobs, and in-process workers are different paths; Moltis also supports remote SSH execution; ZeptoClaw includes a native runtime alongside stronger confinement options.^{[41][47][52][49][46]}

PicoClaw and OpenFang explicitly disclose pre-1.0 maturity limitations. OpenFang's “Hands” package scheduled roles and their tools; the term “agent operating system” does not make it a new host kernel. TinyAGI uses persistent teams and a durable queue but labels itself experimental. These are reasons to scope a pilot and test recovery, not to infer abandonment from a quiet week.^[43][44][48]

BabyClaw is a deliberately small Telegram/VPS application built around Claude Agent SDK. HermitClaw emphasizes continuous autonomous research and reflection. The latter's own warning is unusually useful: its folder restrictions and Python guardrails are not a security boundary. TrustClaw is a connected-services application with external infrastructure dependencies, so operating its app yourself is not equivalent to operating every service it uses.^[55][53][50]

---

Security: Compare the Actual Action Path

“Strong security” is not a useful table value without identifying the process, resource, credential, and enforcement mechanism involved. This comparison does not assign an overall winner from product descriptions.

For any shortlist, inspect filesystem mounts, credentials, existing browser sessions, outbound destinations, confirmation rules, and what survives a restart. Test the exact tools you plan to enable. A secure code runner does not by itself prevent an authorized email connector from sending the wrong message.

Self-hosting changes who controls deployment. It does not remove the model provider, chat service, OAuth connection, or remote-tool operator from the data path. Fully local inference is possible for some configurations, but it must be selected and tested; “your data never leaves your machine” is not an appropriate category-wide promise.^[3][45][50]

---

Resource Requirements and Cost

Compare the same workload

A small executable, a small resident process, firmware size, model weights, and a complete browser/container installation measure different things. They should not share a “smallest agent” ranking without a common workload and measurement method.

• Go-based PicoClaw and Zig-based NullClaw provide compiled runtimes, but the model and any browser/tool infrastructure remain separate resources.^[43][47]
• MimiClaw's documented ESP32-S3 requirements are flash and PSRAM requirements. Its language-model requests still use an external service; zclaw can likewise use a cloud or reachable local model endpoint.^[60][61]
• A Docker desktop workbench like Agent Zero includes applications and persistent user storage beyond the agent process. A messaging gateway and a LibreOffice/browser task are not comparable memory workloads.^[65]
• MicroClaw's SQLite state and delivery queue support persistence; state durability still needs backup and restore testing in the operator's deployment.^[51]

Measure idle memory, memory during the intended task, cold-start delay, storage growth, model tokens, repeated retries, and recovery after interruption. Report the operating system, hardware, model, tools, and version alongside any result.

Subscription details that affect a decision

Manus's help center currently publishes Pro starting at $20/month and distinguishes subscription credits from Cloud Computer capacity. The computer plans start at $10/month and rise with memory/storage; cancelling that add-on affects the computer's persistence. The public pricing route did not render during this review, so these are published help-center terms rather than an independently verified checkout quote.^[5][6]

Lindy's monthly seats include 3,000, 15,000, or 35,000 credits at the Plus, Pro, and Max tiers. Credits pool across a team and unused monthly credits expire. Ako's free 2,000 credits are a one-time starter allowance, not a recurring free allocation; its paid monthly credits expire, while purchased top-ups have separate terms.^[8][18]

HappyCapy's annual equivalents of $17, $42, and $167 per month are different from paying $20, $50, or $200 month to month. Lazzy's reviewed monthly page displays a promotional $5 Pro price with metered credits, not the old $20 figure as the only current option. Plan-specific allowances still matter more than the displayed entry price.^[16][36]

Muse publishes 500 million weekly Muse tokens for Power and 3 billion for Maximum. These are product allowances, not a verified number of tasks or a standardized API-token equivalence. Benefits vary by account/region and paid plans renew monthly; the help page still describes limited testing.^[14]

Perplexity's individual page lists Pro at $20/month and Max at $200/month. The displayed 4,000 Pro and 35,000 Max bonus credits are distinct from Max's 10,000 monthly Computer credits. Portable's local execution changes credit consumption, but it still requires an eligible subscription and suitable hardware.^[33][32]

Nebula's rendered pricing page showed no selectable base plans during this review. It did separately list agent/tool credits and cloud device usage. This report therefore does not repeat an old subscription range as current pricing. Blockit's reviewed site likewise established a trial but not a current paid quote.^[20][25]

Moltworker requires a paid Workers account plus the relevant container, model, and optional storage costs. The README's $5 Workers minimum is not an all-inclusive agent budget. Without optional R2 persistence, configuration and history can be lost when the container restarts.^[59]

A useful pilot tracks cost per reviewed result, including subscription allocation, usage, compute, connected services, and human correction. Scheduled agents can spend money while their owner is not looking; a cheap runtime does not make repeated model calls free.

---

Status, Migration, and Adjacent Tools

These links preserve useful context without adding members to the 45-platform matrix.

When the job belongs to a team

Groovy, Kortix, and Tembo are adjacent because their current organizing unit is an agent workflow, shared project, or team software process. They are considered under the same fit rule rather than counted selectively as personal assistants.

Groovy currently orchestrates local Claude Code/Codex work with shared context, repositories, connectors, and review. Its personal license is $49.99/year for noncommercial use, and enterprise use requires different terms. It is described as source-available; that should not be silently rewritten as an unrestricted open-source license.^[84]

Kortix's current README organizes agents, skills, company memory, and connectors in a versioned repository. Sessions use cloud computers and produce reviewed change requests. It supports managed and self-hosted deployment, but its emphasis is a company-wide agent management system; the older “Suna personal assistant” framing is insufficient for current inclusion.^[85]

Tembo is relevant when a personal task turns into repository work that teammates need to observe, review, and operate. Its platform runs coding agents with repositories, tickets, and connected tools in cloud VMs, returns pull requests or artifacts, supports scheduled/event-driven workflows, and offers self-hosting. This is a meaningful complementary option for engineering work, without making it a replacement for every consumer inbox or household assistant. Disclosure: Ry Walker is Tembo's founder and CEO.^[86]

Tembo Agent Studio is the separate project to consider when an inbox or task assistant becomes a shared operational workflow. Its documented examples cover email triage, ticket roundups, CRM follow-ups, and daily task lists. Agents stage results in a shared Tasks Inbox; a reviewer can edit a proposed reply or choose an action that executes through their connected account. This provides a concrete alternative to leaving every result in an individual chat stream.^[87][88]

Grok Bot's sharing mechanism creates a copy in the recipient's account, without the original computer or logins. Studio's shared review queue is therefore a distinct workflow to evaluate when responsibility crosses from one person to a team.^[89][88]

That overlap makes Studio relevant to personal-agent buyers, but its organizing unit is a team workspace with shared review and managed agent definitions. It remains adjacent to the 45 counted personal-agent members, with fuller treatment in team-agent platforms. The affiliation disclosure above applies to both Tembo products.

NemoClaw and Moltworker remain in the counted operator section because they provide deployable personal-agent environments, with their alpha/experimental limitations explicit. TinyAGI, BabyClaw, HermitClaw, and droidclaw remain where their current implementation fits. None is declared dead from a commit-date cutoff.

---

Decision Guide

Choose by the resource the task needs

These are editorial shortlists drawn from the cited capabilities above, not rankings from a controlled benchmark. A product's absence from a shortlist is not a claim that it cannot perform the task.

Run a representative pilot

Start with one recurring workflow and an explicit output: for example, prepare a weekly meeting brief from approved messages and a calendar, with source links and a draft document. Give the agent a bounded set of accounts and files. Define which changes require review before enabling unattended runs.

Evaluate the entire loop:

1. Context: Does it find the right records, preserve source provenance, and distinguish current instructions from old memories?
2. Actions: Does it use the intended account, avoid duplicates, and stop at the agreed review boundary?
3. Continuity: What happens after a disconnect, restart, expired token, or missed schedule?
4. Recovery: Can the user inspect changes, restore files or state, and correct a bad memory?
5. Cost: What did the accepted result consume, including retries and human correction?

First-hand reports make these checks more concrete. On September 12, OpenAI Community user evilintala reported Work context loss and changes beyond a requested UI fix. Agent Zero GitHub user sumer-man reported early-v2 history growth on June 29; maintainer 3clyp50 described the reproduced cause and a planned fix on July 14. These are specific experiences, not category-wide failure rates or proof of current defects in every release.^[90][91]

---

Market Outlook

The supported direction is broader execution choice: consumer services now expose cloud computers and local application access, while open implementations offer more runtime and memory options. Work's local/cloud split, Cowork's cloud default, Letta's persistent agents, and Agent Zero's explicit v3 isolation proposal illustrate different ways to support longer tasks. A proposal remains a proposal until a shipped version establishes the behavior.^{[64][2][21][57]}

The next useful comparisons will measure successful repeated work, recoverability, permission enforcement, and accepted-output cost. Raw stars, a vendor's smallest-binary claim, or a forecasted market share do not answer those questions. More integrations can expand usefulness and the number of accounts an operator must understand.

The likely buying split is between people who want the provider to operate the environment and people who want to customize and maintain it themselves. That is an editorial outlook, not a numerical forecast. Both groups still need a way to inspect what an agent did and decide whether the result is good enough.

---

Bottom Line

Choose the agent around the task's data, tools, execution location, and recovery needs. Hosted products reduce infrastructure work; self-operated projects increase configuration control and responsibility. Device agents can reach resources a cloud service cannot directly access, while cloud tasks can continue when a laptop is off.

The strongest starting point is a small, repeatable workflow with measurable output and a clear review boundary. Expand access only after the specific deployment demonstrates that it can complete that work, preserve context, recover from interruption, and stay within an acceptable cost.

For a companion overview of self-operated options, see 6 OpenClaw Alternatives Worth Considering in 2026.

---

Research by Ry Walker Research • methodology