karta.toml: per-agent commands resolve an explicit --agent or slug first,
then the agent in this folder, the repo’s sole local agent, and finally the
legacy KARTA_PROJECT alias. Deployment itself is folder-scoped even when
several agents share one git repository.
Deploy
karta deploy ships one agent project using the deploy_method in its local
karta.toml. On first git deploy it provisions the hosted source and wires the
repository’s karta remote; folder deploy uploads only the selected folder.
For low-level git use, target the agent’s reserved deploy ref explicitly:
git push karta main is a compatibility convenience for a repository
with one enabled agent. It never fans out a multi-agent repository. See
Releases & rollback for the deploy flow in
depth, including plain folders and CI.
Inspect
Roll back
karta rollback when you mean “revert production.” Use
karta agent activate when you want to activate a specific existing production
release as an explicit operation.
Agent controls
Logs
Open
karta.toml; elsewhere, pass --agent <slug> or set KARTA_PROJECT in CI.
Clone a karta’s workspace
Each karta is a durable computer, and its files persist across sessions. With workspace access enabled, an org admin can clone that workspace read-only over git to see what the agent has accumulated for a karta. Turn it on in the console first: on the agent’s Settings tab, enable Workspace access, or run:karta login installs a git credential helper that authenticates the clone for
you - no token to copy or paste. Access is read-only: clone and pull
work, push does not.
Where sessions fit
For local Karta parity, usekarta dev: it runs your agent behind the
same session API on your machine, with hot reload and a REPL - see the
CLI overview. In production, sessions are
driven through the session API by your frontend or
the chat widget; the hosted chat page (karta open)
is the quickest way to talk to a live deploy.
