Kavela vs ChatGPT's cloud agents

1. Your own model plan

How ChatGPT does it

OpenAI's dots is powered by GPT-6 Astra, and ChatGPT Work's model picker lists OpenAI's GPT-6 models ^{2, 7, 15}. In Codex Cloud, the models on offer are the hosted Codex models your ChatGPT sign-in gives you access to, and OpenAI's Amazon Bedrock setup is described for local ChatGPT Work and Codex surfaces ^{16, 17}. None of the three describes a way to use another company's model or subscription ^{2, 7, 16}.

Your first dot is included in a Pro or Business Premium plan at no extra cost, and Pro costs $100, $200 or $500 a month ^{1, 18}. On Enterprise, dots is off by default ². Conversations with the dot sit outside ChatGPT usage limits; the Work and Codex tasks it starts count against them, and Work and Codex share one allowance ^{1, 2, 18}. OpenAI does let selected partner apps use a person's ChatGPT plan through Sign in with ChatGPT, which is in a limited trial for commercial partners ¹⁹.

How Kavela does it

Kavela runs on your choice of plan, agent by agent. With your own ChatGPT plan, Codex runs inside the agent's own cloud computer and you sign in to it yourself; Kavela does not route requests through anyone's plan itself. With a Claude plan, the same agent can run Claude Code instead, and Kavela's included Kav AI usage covers anyone who has neither.

Because the choice is per agent, one team can mix Codex and Claude agents, and you keep paying for the subscriptions you already have.

You keep the ChatGPT plan you chose, and it can work next to a Claude plan on the same team.

2. Managing a team of agents

How ChatGPT does it

OpenAI's dots coordinates other agents. It can divide work among background agents that run in parallel and report back to it, create separate threads you can see and send follow-up instructions to, and start ChatGPT Work or Codex tasks on a connected computer or in a Codex cloud environment ⁵. It decides for itself when to pause and wake up to continue, so not every follow-up needs a fixed schedule ².

Today each person starts with one primary dot; OpenAI describes teams of dots working together as a later goal ¹. You follow the work in lists: delegated tasks under Activity, recurring work under Scheduled, and cloud threads as separate conversations in the ChatGPT apps ^{2, 6}. The docs describe no calendar view and no standing team of named workers; the helpers they describe are background agents that run in parallel and report back ^{5, 6}.

Inside ChatGPT Work, one task can split independent work across subagents and follow their progress, but the docs describe no coordinator across separate tasks ¹². Codex Cloud runs the tasks you start in parallel, each with its own working state ⁴.

How Kavela does it

In Kavela the lead agent is Kav. You tell Kav what you want, in the workspace chat or in Telegram, and Kav plans it, hands each part to the right worker agent, follows up and brings back the result. The workers are a standing team: named agents with their own skills, connected apps and run history, and you can hire a new one by describing the job.

Every task, from every agent, sits on one calendar, by day, week or month, with its start time, due time and live state, so a week of agent work reads like a week of a team's work: what is scheduled, what is running, what needs you and what is done.

Both have a lead agent. Kavela's leads a standing team of named agents, and a week of their work reads on a calendar instead of a list.

3. Agents that can sign up for things

How ChatGPT does it

The dots docs describe no email address of the agent's own. Connecting a messaging channel does not connect your inbox; your dot reads email through plugins such as Gmail ². ChatGPT Work's docs describe none either: Work uses your own Gmail through a plugin and can start a task when a new message arrives ^{9, 15}. For enterprises, OpenAI is piloting specialist dots with their own identity for access management ¹.

OpenAI is strong on the step that usually follows: signing in. When a website asks a dot to sign in, you enter your credentials into a private form that goes to the remote browser, outside the conversation, and complete any verification the site requires; you can save the password for later, and reusing it needs your confirmation ⁸. ChatGPT Work signs in to websites through a secure form too, on the web and mobile, on Plus and Pro ²⁰.

How Kavela does it

Every Kavela workspace has its own receive-only inbox, Kavela Mail. When an agent signs up for a service on your behalf, it uses a fresh address from that inbox, waits for the verification email, reads the code or the link and carries on. The wait shows in Approvals, and received mail is deleted after 30 days.

Signing in to an account you already have works much as it does in dots: you sign in through a live browser you control, and the password goes into the card, never through the model. The inbox covers the other case, an account that does not exist yet: the agent can create it, read the verification email and finish setting the service up.

OpenAI's agents sign in with the accounts you already have; Kavela agents can also create the ones a job still needs.

4. What agents need, and what they deliver

How ChatGPT does it

Before an action that affects your accounts or shares information, dots runs an automatic review that decides whether it can go ahead, needs your approval or must hand the step to you. You can add your own rules with four settings: take action without asking, take action when you say so, ask before taking action, or hand off to you ^{2, 6}. When a task is waiting for a decision, an app connection, a sign-in or an approval, you open the task in Activity and answer the request there ⁶.

ChatGPT Work pauses in the conversation when it needs information or approval, and always asks before consequential actions such as booking an appointment or completing a payment ^{12, 20}. On the desktop, an Activity view lists chats that are unread, running or waiting for your response ²¹. Work delivers files you review and download, from documents, presentations and spreadsheets to PDFs, as well as Sites; Codex Cloud delivers changed files and test results that you commit or open as a pull request ^{4, 12}.

The docs describe no single inbox that gathers every pending request across tasks as typed cards; each request stays with the task or chat that raised it ^{6, 21}.

How Kavela does it

Kavela collects everything your agents need into one Approvals inbox for the whole workspace. Each request is a typed card you answer where it is: connect an app, sign in to a site in a live browser, approve a post before it goes out, choose between options, or see that an agent is waiting for an email. Other work carries on while a card waits.

What agents produce arrives as deliverables attached to the task that made them, from documents and reports to sites, listed under Delivered and opened in one click.

OpenAI's rules decide when an agent asks you. Kavela also settles where you answer: one inbox, for every agent.

5. Many agents, running on schedule

How ChatGPT does it

Each dot has its own computer and browser in the cloud, and the computer can keep its state between periods of use ⁸. Recurring work needs a saved schedule that sets what to do, when and in which time zone, which changes to report and where to deliver them, and a dot can watch for events when a connected service supports it ⁵.

ChatGPT Work's cloud tasks run in an isolated hosted environment and keep going after you close the desktop app ¹². Scheduled tasks on the web can use uploaded context and connected tools; event triggers on Gmail, Slack and GitHub run on eligible plans, and one task cannot combine an event trigger with a time-based schedule ⁹. Codex Cloud gives each new task its own isolated workspace, with up to 4 vCPUs, 16 GiB of memory and 32 GiB of disk on Pro, Business and Enterprise, and keeps a task's saved state recoverable for up to seven days ¹³.

How Kavela does it

Each Kavela agent works in its own cloud computer, with a shell, files, git and a browser, which starts when work starts and sleeps when it stops. One agent's work never slows another's, and work keeps going after you close the tab. The workspace's files persist between runs, so tomorrow's run starts from today's output.

Automations run on a schedule or on an event, such as a webhook or a merged pull request, and can be limited to an active window with start and end dates. There is no cap on how many are switched on at once.

Recurring work belongs to the whole team: every agent keeps its own computer, and each run starts from the files the last one left.

6. Connecting your tools

How ChatGPT does it

OpenAI's agents reach a lot of apps. A dot connects to over 4,000 apps through plugins, and your existing ChatGPT app permissions apply ^{1, 8}. In hosted ChatGPT Work you install a plugin to use its bundled connectors and remote MCP tools; the docs describe adding a new MCP server by packaging it as a plugin, and no way to paste in a server's address ^{22, 23}.

For websites, dots and ChatGPT Work sign in through the private form described above, and a dot can also use your own computer, one at a time ^{2, 8, 20}. Work's website sign-in is not available in Enterprise or Edu workspaces, and some sites block cloud browsers or ask for extra verification ^{2, 20}.

Codex Cloud adds options built for developers: environment variables, network secrets that programs see only as a placeholder, a personal vault, a Tailscale VPN and, for Enterprise on request, OIDC ¹³. Its environments are built from GitHub repositories, and computer and browser use are listed among its current limitations ^{4, 13}. The dots docs describe no API keys or custom secrets of the dot's own ⁸.

How Kavela does it

Kavela connections belong to the workspace: connect a service once and every agent can use it. A service can be connected by OAuth sign-in, as an MCP server, with an API key, with custom secrets, or by signing in to it in a live browser you control. Passwords and keys go into the card, never through the model.

You can also leave it to the agent. With Let Kav set it up, the agent reads the service's documentation, works out the right way to connect it and does the setup, creating the account first if there is none.

A large plugin library covers the common apps; an agent that can connect a service from its documentation covers the rest.