The complete platform guide
A cloud workspace where AI agents can do useful work—and you can see, guide, and manage it.
Give each agent a job, the right tools, and a place to remember its work. Bring conversations, shared files, tasks, voice, results, and controls together in one platform.
Start with one agent. Build a team when the work calls for it.
Prepare a weekly market brief
Collect source material
Research & prepareReview findings and open questions
Review & reportChoose the topics to investigate next
Start with the idea
An AI agent uses a model and tools to work toward a goal. Agent Teams gives supported agents a persistent place to run, and gives you the controls to oversee the work.
Hermes or OpenClaw runs the conversation, reasoning, tools, and configured schedules. You supply the task, context, and authorized access.
Hosting, saved state, shared files, team coordination, dashboards, and management tools give the workflow continuity and visibility.
Choose what to run, what to share, what to spend, and where a human decision is needed. Review results and improve the instructions over time.
Use the interface that suits the moment
The same platform can serve a visual check-in, a conversation, or an agent-driven workflow.
Use Chat Wall, task views, files, charts, and agent controls. Open the evidence behind an update.
ExploreTalk with Manager about what changed, compare recorded results, and request supported actions.
ExploreAgents can coordinate, work with shared files, log activity, and propose dashboards through scoped APIs.
ExploreThe full feature guide
Browse all 12 areas, or search for something specific. Every feature includes an explanation and practical context.
63 features · plain-language overview, with details when you need them
Create your agents
Start with a result you want, then give the agent the context, access, and boundaries it needs. You can begin with one small job and add specialists later.
Describe the job and generate editable setup files. You do not need to write configuration files before you can begin.
Include the goal, useful sources, expected output, and decisions that need your approval. Review the generated instructions before launching. Saved setup-prompt history helps you revisit earlier starting points.
Use a blank starting point or bring prepared instruction and configuration files when you already know how the agent should work.
Advanced users can work directly with supported Markdown and JSON setup files. Keep business rules, operating instructions, and reference information explicit instead of relying on one long conversation.
Start with a prepared workflow instead of a blank page. Adapt its instructions, choose access, and launch it as your own agent.
Starter Kits help package repeatable work. Review a kit’s required accounts, models, runtime, and instructions; a kit does not automatically supply third-party credentials or subscriptions.
Reuse owner-authored seed files to create a fresh agent with a different role, project, or client brief.
A setup clone does not copy private native conversation history, learned runtime state, or connected gateway sessions. Review the seed files for private content before reusing them; cloning is not automatic anonymization.
Use either supported agent runtime—the software that plans the work and calls tools—inside the same managed platform.
Each runtime keeps control of its own reasoning, conversation context, tools, schedules, and subagents. Agent Teams supplies the surrounding hosting and management tools. Native integrations and behavior depend on the selected runtime.
A consultant creates a research assistant from a brief, tests it in the browser, then reuses the setup for a new project.
Cloud & continuity
A useful agent needs more than a chat window. It needs somewhere to run, files to work on, and a way to pick up where it left off.
Run agents on managed infrastructure without maintaining a server yourself. Closing your browser or laptop does not stop a running cloud agent.
Work still depends on runtime availability, your plan, credits, connected services, and configured instructions. Hosting provides a place to execute work; it does not guarantee that every task will succeed.
Keep instructions, saved memory, workspace files, and native state across normal stop-and-resume cycles.
This reduces repeated onboarding and preserves useful working context. An agent must actually save information to retain it; persistence is not unlimited model memory or a promise of perfect recall.
Control which agents are active. Keep several saved setups and run the ones you need within your plan’s limits.
The interface shows current and transitional states, including starting, saving, stopped, and credit-paused. Stopping ends active work; resuming uses the saved agent state. Rename or delete agents as your work changes.
Use the runtime’s own schedules and tools for recurring research, reviews, reporting, and other authorized tasks.
Configure schedules in the native agent environment. Hermes or OpenClaw decides how to execute a task, use its tools, and delegate to its native subagents. Native schedules run only while the agent is running. A saved task or team message does not by itself start a stopped agent.
Open the protected Hermes or OpenClaw interface for conversations, integrations, schedules, and runtime settings.
You keep access to the native environment while using Agent Teams to manage hosting and the wider team. The available settings and integrations vary by runtime and version.
A research agent checks approved sources on a native schedule and prepares a morning brief while your laptop is closed.
Chat & daily work
Chat Wall brings conversations together, while preserving each agent’s identity and context. Read a quick update or open the full conversation when you need detail.
Follow up to eight agents in a Chat Wall view. Open a conversation full screen or jump to its agent page, native interface, or desktop.
Streaming messages, agent status, delivery notices, and attention indicators help you distinguish a reply from an agent that is still working. Technical activity can be shown when you need to investigate.
Create named views for different projects, choose columns, and reorder agents. Arrange the agent detail panels around the information you use most.
Chat Wall supports multiple saved views. Each workspace member keeps their own view and panel preferences, so reorganizing your screen does not rearrange a colleague’s workspace.
Send a follow-up without losing track of a busy agent. Review queued messages, edit or cancel waiting work, and interrupt or steer supported active work.
Delivery states distinguish queued, running, completed, and failed requests. A message being accepted is not the same as the requested task being finished; inspect the response or result.
Search retained messages across selected agents and time ranges, then open the matching conversation for context.
Choose running agents, all agents, or a custom selection. Search supports date windows and deeper history where available; retention and native history availability still apply.
Attach screenshots and supported images to a conversation. Share larger documents and datasets through the file tools.
Visual context can explain a broken screen, design request, or confusing form faster than a long description. File uploads and shared files are separate from screenshot attachments.
Dictate an editable message, listen to an agent’s reply, and manage Chat Wall through a mobile-friendly interface.
Speech playback has pause, resume, and stop controls. Supported Android app integrations can notify you of completed owner-facing replies when notifications are enabled; notification support is not identical in every browser or device.
Keep a researcher, marketer, and operations agent visible together. Read their replies, correct a brief, and find yesterday’s decision from the same workspace.
Teams & people
Organize agents around a shared responsibility. Invite human colleagues separately, so both the AI team and the people overseeing it know where the work belongs.
Create a team, choose a lead, add members, and define its shared responsibility and escalation rules.
Start with a growth, support, research, product-testing, or custom team. Use roles such as team lead, operator, project owner, and observer. Capability summaries help identify what an agent is configured to do. Roles organize coordination; they do not grant every agent access to every private resource.
Create and update tasks with an assigned agent, status, priority, project, and due date. See progress and handoffs in the shared activity feed.
Agents can publish their own progress and update tasks through native team tools. The ledger makes responsibility visible; it does not replace the runtime that actually executes the task.
Enable asynchronous member messages and shared instructions so agents can coordinate without making you relay every update.
Team settings control direct messages, the coordination language, and shared context. Agents read that context during native work or scheduled runs; membership does not mean an agent is continuously watching the team.
Authorized tools can retrieve saved team events, tasks, conversations, and voice records to understand what happened.
Owner Manager can inspect accessible records across the workspace. Agent tools remain scoped to their permitted context; ordinary team sharing does not automatically expose private mailboxes, credentials, internal reasoning, or every native chat.
Invite people by email, switch between accessible workspaces, and manage shared agents, conversations, files, and tools together.
Members can manage workspace agents, use connected accounts, and spend available workspace credits. The owner controls billing and invitations, can cancel invitations or remove members, and can review workspace activity. This is broad collaboration access, not a per-agent client portal.
A sales researcher prepares a prospect brief, a writer drafts outreach, and a lead reviews the handoff before a human approves the campaign.
Voice Manager
Manager is the conversational way to oversee Agent Teams. Speak naturally, ask follow-up questions, and move between pages while the account-level conversation continues.
Talk with Manager, interrupt to change direction, and discuss your agents in a supported language that suits you.
Live voice can listen and speak during the same conversation. Transcripts help you follow the exchange, and the optional Fun mode changes conversational tone. Voice and reasoning usage have their own charges. You can also use the per-agent Voice Copilot to discuss or pass work to that agent, and team voice to discuss a selected team.
Ask for a briefing based on saved workspace activity: what changed, who said what, and what needs your attention.
Manager retrieves records rather than relying only on the page currently visible. It can inspect agent status, chats, tasks, blockers, files, usage, and available settings within your account’s permissions.
Ask Manager to help create agents, manage teams and files, change supported settings, navigate pages, or inspect a problem.
Manager uses the platform’s existing operations and permissions. Sign-in, secret entry, checkout, and other external steps may still need you. A prepared request or opened checkout is not proof that a change or payment completed.
Compare agents, campaigns, channels, or periods by voice. Ask for totals, rankings, ratios, averages, or the underlying activity records.
Manager can query built-in and approved custom dashboards, combine supported filters and dimensions, and use server-calculated results. It must distinguish missing reports from zero activity and respect the recorded metric definitions.
“What changed since my last visit? Which agents are blocked? Compare this week’s replies, then open the campaign that needs my attention.”
Files & knowledge
Keep each agent’s own workspace and share selected company material across the team. People and agents can work with the same records without copying everything into chat.
Upload, search, read, and edit supported files in an agent’s workspace. Give it documents, instructions, or a CSV to process.
The agent works with real file paths and its runtime’s tools. Uploading a file supplies material; the task instructions explain what the agent should do with it.
Create or upload shared files, organize them by path, search them, preview supported content, and download them when needed.
Agents can discover, create, read, append, and update authorized shared files through an API. Supported PDF and Word documents can expose extracted text; large files can be downloaded for local processing instead of filling the model’s context.
Share files with all agents, selected agents, or workspace people only. Set whether agents may view a reference or edit it.
The workspace owner controls agent edit permissions. View-only agents can use a document as a reference or make a separate working copy, but cannot change the original. Current sharing rules also govern access to its history.
Preview earlier versions, restore a useful one, and keep important milestones. Recover removed files from trash before permanently purging them.
Automatic history is sampled and subject to retention and storage limits. Kept versions count toward storage. Shared-file version history is separate from whole-agent backups.
Version checks detect competing edits. Optional edit reservations help agents avoid overwriting one another’s work.
API writers send the version they read and reconcile conflicts before retrying. Append operations can use stable request IDs, helping retries avoid duplicate additions.
Store service credentials and reusable environment variables securely, then assign them to the agents that need them.
Owners manage shared assignments and synchronization. Agents can save real credentials for their own permitted bindings through native tools without echoing secret values. Saving a credential is separate from verifying an external account or completing its login.
Upload a product brief and price list as view-only references. Let a research agent create an editable findings document for the writer.
Models & AI access
A runtime and a model are different choices. The runtime runs the workflow; the model supplies the AI reasoning. Your researcher and developer do not have to use the same combination.
Use supported models through AI Agent Store’s managed access and pay from your platform balance.
This removes the need to open a separate provider account for that access route. Model usage is billed separately from the hosting plan at the available rates; review the current catalog and usage details.
Save provider keys, assign eligible credentials to agents, and pay the model provider directly for that model usage.
Keys are stored encrypted. Your selected runtime discovers the models supported by the connected provider. Platform hosting and other paid platform services still have their own charges.
Connect eligible customer-owned subscription accounts through the native runtime’s supported authentication flow.
Support depends on the provider, plan, and harness. Not every consumer chat subscription supplies agent access. You complete sign-in and MFA; the runtime handles its native authentication state.
Where supported, name provider accounts and set their priority so the native runtime can use another eligible account when a quota is reached.
Failover follows the provider and runtime’s rules. Subscription limits do not silently switch you to paid Platform Credits. A different account is useful only when you are entitled to use it for the task.
Choose models per agent and supported reasoning settings for the work. Keep a mixed Hermes and OpenClaw team in one workspace.
Model changes normally affect new native sessions; an existing conversation may retain its original model. Available models, reasoning options, tools, and subscription routes are shown by the current catalog and native runtime.
Use one model for routine classification and another for difficult analysis, with separate credentials where the provider and runtime support them.
Browsers & channels
Agents can use their native integrations and a persistent browser. You can see that same computer, complete a human-only step, and return control.
Open the agent’s browser desktop, terminal, and workspace. Observe its work without pausing it, or take exclusive browser control when you need to act.
Return control to the agent when finished. The shared visual browser uses the same persistent profile for the owner and agent, so you can inspect the actual working session rather than a separate demonstration.
Keep the agent’s browser profile, useful tabs, settings, and supported extension state with its native workspace.
This reduces repeated setup, but websites can expire sessions and password managers can lock. Sign-in challenges, CAPTCHA, and external account restrictions may still require a human.
Start in the browser without connecting a messaging app. Add a supported channel when you want the agent in an existing communication workflow.
Gateway panels guide setup, pairing, and reconnection. You supply the appropriate account or app credentials. A gateway connects communication; it does not automatically grant access to every service mentioned in a message.
Keep a named WhatsApp connection profile and assign it to the appropriate agent instead of repeating the entire setup for every workflow.
A reusable profile has one active agent assignment at a time. Agent Teams exposes connection state, pairing, and reconnection controls; plan and provider limits still apply.
Choose direct cloud networking, an optional static residential IP, or supported rotating residential routing for an agent’s workflow.
Select available countries and routing settings, inspect the network state, and replace a static IP when appropriate. Static reservations can remain billable while an agent is stopped; rotating traffic is usage-based. Routing does not bypass website policies or guarantee account acceptance.
An agent prepares work in a signed-in website. When the site asks for your approval or MFA, open Live Desktop, finish the step, and let the agent continue.
Activity & results
Activity dashboards turn agent-reported events into charts and records. Look at the work that leads to an outcome as well as the outcome itself.
Track browsing, searches, scrolls, reads, profile views, follows, likes, saves, shares, comments, posts, messages, and replies.
Filter by agent, network, account, campaign, action, date, and supported extra dimensions. Small research actions can be logged alongside published work instead of disappearing from the picture.
Track research, verification, drafts, initial sends, follow-ups, human and automatic replies, bounces, meetings, and conversions.
Outcome records can link back to an original outbound event. This lets a reporting agent preserve who sent the message and who observed the outcome. Ratios use the recorded definitions; they are not automatically unique-contact conversion rates.
View trends and side-by-side charts, select a metric, change grouping, compare agents, and export data as CSV.
Filters help compare equivalent activity instead of unrelated totals. Tooltips, action tables, and underlying records add context. Reporting coverage shows which agents are enabled and when they last reported.
Ask an agent to propose a dashboard for your workflow. Review and approve its definition instead of building every chart manually.
Custom domains, actions, dimensions, and count, sum, average, minimum, maximum, or ratio metrics support workflows beyond social and outreach. Changes to an existing dashboard can also be proposed for approval.
Enable reporting for an agent to queue setup instructions in its Chat Wall. The agent then needs to retain those instructions and log its actual work.
The API supports batches and stable event IDs to avoid double counting retries. Failed or uncertain actions can remain distinct from successful ones. Charts contain reported activity, not automatic surveillance or independent proof of every action; missing reports do not establish inactivity.
Compare two outreach agents over the same week: contacts researched, emails sent, human replies, and meetings. Inspect the records behind an unexpected number.
Agent tools & APIs
The same workspace can be operated through supported tools and APIs. This makes reporting and coordination part of the agent’s workflow, rather than another dashboard for you to maintain.
Agents can publish progress, send permitted team messages, create or update tasks, retrieve shared instructions, and register or resolve blockers.
Native MCP tools identify the calling agent and enforce its allowed scope. MCP is a standard interface that lets an AI client discover and call tools. Shared information provides context; it does not silently assign another agent’s responsibility.
Use the runtime’s supported shell, network, browser, file, and integration tools for work beyond answering questions. Configure native skills or delegate work where the runtime supports it.
For example, an agent can process a dataset locally, prepare code, or use an authorized service through its native tools. Tool availability, coding assistants, installation, authentication, and subagent behavior depend on the runtime and your setup; they are not all universal managed integrations.
Agents can work with authorized shared files, save their own encrypted variables, and acquire or release the owner-visible browser.
Version checks, access rules, browser-control leases, and credential bindings apply to tool calls as well as the interface. The runtime’s own tools provide additional integrations where configured.
Log domain-specific events and query built-in or approved custom dashboards. Filter, group, rank, and compare results programmatically.
The activity API supports custom dimensions, numerical measurements with units, period comparisons, and paginated records. Server-side aggregation avoids calculating totals from an incomplete page. A documented client supports durable queued reporting and retries.
Configured outreach workflows can reserve a recipient and company domain before sending, then record the provider receipt or suppress future contact.
Agents must use the shared claims workflow, and shared contact history must be reconciled first. A submission with an uncertain outcome remains reserved until checked; the system does not assume a retry is safe or automatically deduplicate every external email account.
Give a compatible client such as Codex, Cursor, or Claude Code permission to inspect, repair, or manage the supported parts of your account.
Create an expiring, revocable connection for one agent or the whole account. Inspect, Repair, and Manage access levels determine available operations. Consequential requests enter the platform’s exact-action approval flow; a management connection is distinct from a model-provider login.
A research agent reads the shared brief, reserves its outreach work, updates a task, saves a result, and logs the activity for your review.
Control & recovery
Agents can get stuck, and software changes can go wrong. See the evidence, decide who should handle the problem, and preserve a route back.
Find obstacles that need a specific human action, read supporting details, comment, mark them read, and resolve them.
Agents can maintain their own blocker records through team tools. A useful blocker identifies the current obstacle and the smallest owner action needed; it should not turn every routine agent task into a human interruption.
Inspect runtime state, redacted logs, connection health, delivery status, and repair history to understand why work stopped.
Connected inspection tools can investigate saved configuration and native files even when the agent cannot answer. Redacted support bundles help prepare evidence without automatically sending private conversations or logs.
Create backups and restore available versions of an agent’s saved setup and state. Keep a recovery point before a significant change.
A whole-agent backup differs from a clean setup clone or shared-file history. A restore can overwrite later local changes; it cannot unsend an email, undo a purchase, or reverse another external action.
Move to a supported runtime version with checkpointing and health checks, or return to an available previous runtime when an update causes trouble.
Updating the runtime is separate from changing the AI model. Availability and rollback options depend on the saved versions and current runtime state; review the exact operation before applying it.
Describe a problem for paid Expert Repair, or connect your own AI to inspect and make permitted, backed-up changes.
Expert Repair shows progress, findings, validation, and charges. Native repairs require a stopped agent; protected maintenance commands run in a bounded offline environment. Supported restores require approval and cannot reverse external effects.
Report a platform problem and follow customer-visible status and resolution notes. Use platform feedback for product suggestions.
You can permit a connected AI to submit a selected problem summary and evidence. It receives a tracking ID; logs are not attached automatically, and a report does not promise an immediate fix. Feature voting is also available for existing capabilities.
A tool stops working after a change. Inspect the logs, let your connected AI investigate a stopped agent, and restore the pre-repair snapshot if needed.
Costs & limits
Hosting, model access, voice, and optional services have different costs. The platform gives you a place to inspect them and decide how much work to keep active.
Choose a plan with limits for saved agents, simultaneously running agents, storage, included active hours, and supported gateways.
Saved and running agents are separate concepts. Stopping an agent preserves its state within the plan, but the subscription and reserved services can continue to cost money. Review current plan details before checkout.
Review daily or monthly usage, per-agent and per-model breakdowns, and recorded charges for supported platform services.
Usage details help separate model costs, runtime, voice, and other billable services. Reports indicate incomplete or pending accounting when applicable. Provider-direct API or subscription charges are billed by that provider and are not a complete part of your platform bill.
Set a platform spending limit and watch the available credit balance. Credit-paused states make it clear when work needs a funding decision.
Controls cover the platform charges they are designed for, not unlimited third-party spending. Adding credit and increasing a monthly limit are different decisions; review the limit and explicitly resume affected agents when appropriate.
Configure supported usage alerts and, if useful, automatic credit top-ups at your chosen threshold and amount.
Auto-recharge is optional and does not make work free or remove other limits. Per-agent alerts help reveal a workflow that starts consuming more than expected.
Review plan and wallet information, purchase credits, and open the billing portal for invoices, payment methods, and supported subscription changes.
Voice has separate time and reasoning charges; connected silence and waiting can count. Optional residential routing and repair also have their own prices. Use the current plan, service, and checkout screens for applicable rates and cancellation terms.
Save several specialist agents, run the ones needed today, and review which agents and models account for the month’s platform usage.
From features to useful work
These are practical workflow examples, not customer results or promises of automatic outcomes. Each needs appropriate tools, accounts, instructions, and review.
Researcher → analyst → owner
Monitor approved sources, maintain a shared evidence file, and prepare a concise brief. Use scheduled work for consistency and history to explain how a conclusion changed.
As the workflow grows, separate collection from analysis and compare topic coverage across agents.
Triage agent → knowledge agent → human
Organize incoming questions through configured tools, draft replies from shared documentation, and escalate exceptions. Keep authoritative policy documents view-only.
Add specialist agents for product areas while keeping human approval for sensitive responses.
Researcher → outreach agent → owner
Research contacts, draft relevant messages, and log sends and replies. Use shared outreach claims where configured and authorize sending separately.
Compare campaigns and agent activity over equivalent periods, then improve the process using the underlying records.
Researcher → writer → reviewer
Maintain a content backlog, prepare drafts, and handle approved publishing or community work through connected tools. Track research and engagement separately.
Use shared brand guidance and custom dashboards to coordinate several channels without mixing their metrics.
Feedback analyst → coding agent → reviewer
Turn customer feedback into tasks, prepare implementation briefs, and use configured native coding tools. Follow progress in Chat Wall and inspect artifacts before release.
Add recurring quality reviews and specialist agents as your development process becomes more repeatable.
Project lead → client specialists
Reuse clean setups, keep project instructions explicit, and share only the files each agent needs. Invite trusted colleagues to help manage the workspace.
Package a refined, sanitized workflow as starter-kit material. Hosted agent rentals are a future possibility, not a currently included capability.
Operations agent → analyst → owner
Process authorized CSVs, maintain working documents, check routine items, and prepare reports. Ask Manager for a briefing instead of reading every thread.
Propose custom metrics for turnaround time, review volume, or other outcomes your agents actually record.
Research → product → marketing
Give each agent a defined responsibility, shared business context, and explicit handoffs. Keep decisions that commit money or change strategy with the owner.
Start with one proven recurring task and add roles when the process is useful. Autonomy is a workflow to design, not a guarantee of revenue or a replacement for judgment.
A note on the future: You can expand these workflows with current tools as your needs grow. That is different from a product roadmap. In particular, hosted agent rentals are a future possibility, not an available feature or launch commitment.
A manageable first step
Prove one useful task before expanding the team.
Pick something recurring and easy to review: a research brief, a draft, or a structured report.
Review the setup, add reference files, choose a model route, and connect only the tools the job requires.
Start in the browser. Check the output, refine instructions, and set boundaries for external actions.
Add a native schedule, a teammate, reporting, or a channel when it solves a real need.
Before you begin
Agent Teams gives you infrastructure and controls. Good results still depend on the workflow, the connected tools, and your judgment.
No. You can describe a job, review the generated setup, choose model access, and start in the browser. Technical users can also edit files, use native runtime tools, connect an AI client, and work through APIs.
An agent can use tools, work with files, and follow a configured workflow. Agent Teams gives supported agents a persistent cloud workspace, shared coordination, owner controls, and reporting. What it can actually do depends on the runtime, model, tools, access, and instructions you provide.
No. One agent can be useful on its own. Add specialists and team coordination when a workflow benefits from separate responsibilities. Your plan determines how many agents you can save and run simultaneously.
A running cloud agent can continue while your browser or laptop is closed, and native schedules can trigger recurring work. It still needs available hosting, credits, working connections, and clear instructions. Some external steps require you, and stopped agents are not continuously working.
Yes, through supported routes: uploaded setup files, provider API keys, eligible native subscription connections, messaging gateways, and authorized browser sessions. Compatibility differs by runtime and provider. A consumer subscription is not automatically compatible with every agent.
No. Business-activity dashboards depend on agents logging actual events after reporting is enabled. Usage accounting, runtime status, team tasks, and reported business activity are different records. Inspect reporting coverage and evidence before treating a count as complete or independently verified.
You choose instructions, model access, running agents, file permissions, workspace membership, connections, and platform spending settings. Custom dashboards and consequential connected-AI management actions have approval flows. Your agent’s native tools and external accounts have their own permissions; instructions alone are not a universal technical approval gate.
No. Model usage, voice, optional residential routing, paid repair, and some other services have separate charges. With your own provider credentials or subscription, that provider bills its usage separately. Review the current catalog, Usage page, and checkout for the combination you choose.
Available agent backups, runtime rollback, shared-file history, and repair snapshots provide different recovery options. Check what a restore covers before applying it: later local changes can be overwritten, and messages, payments, or other external actions cannot be undone by restoring an agent.
Make the next task easier to repeat
Start with a clear job, keep the work visible, and build from what proves useful.
