How Kavela works
Where agents run, what runs them, the harness around them, and what happens when they need you.
Kavela runs every agent in the cloud, in a computer of its own, inside a harness that gives it instructions, playbooks, memory, your workspace's tools and a reviewer.
Where agents run
Each agent has its own cloud computer.
| Part | Detail |
|---|---|
| CPU and memory | 2 vCPUs, 8 GiB |
| Software | A shell, files and a real browser |
| Isolation | Separate from every other agent's computer |
| Browser sign-ins | Kept between runs. You can take over the browser at any time. |
| Between runs | Sleeps after about 2 minutes with nothing to do. Files are saved first and restored on the next run. |
Nothing runs on your laptop, so a run continues when you close it.
Workspace files
Files belong to the workspace. Every agent and every person in the workspace works from the same files.
Runtimes
Each agent runs on one of two runtimes.
| Runtime | What runs |
|---|---|
| Your Claude or Codex plan | The Claude Code or Codex CLI, inside the agent's computer, signed in with your account. See Connect your Claude or Codex plan. |
| Kav 1.0 | Kavela's own model, included with every plan. |
The harness
On either runtime, every agent gets:
- Instructions written for its job. On your plan, they are the CLI's system prompt.
- Playbooks for the tools it uses, loaded when a task needs them.
- Memory of the work you approved and the work you sent back.
- Tools for the workspace's files, connected apps and approvals. On your plan, the CLI calls them over MCP.
- A reviewer that checks each workflow run against its checklist.
Kav
Kav is the lead agent in every workspace and the one you talk to. Kav plans a job, files each part as a task, assigns an agent to each task and creates a new agent when none fits.
Approvals
When an agent needs a decision, a sign-in or a password, it posts the request to Approvals and ends its turn. Your answer starts its next turn from where it stopped.