Getting started¶
terok-executor runs an AI coding agent inside a hardened, rootless
Podman container — one CLI command to launch, or one Python class to
embed in your own tooling.
Why terok-executor¶
AI coding agents need network access and credentials to do useful work, but giving them uncontrolled access to either is a risk: a prompt-injected or supply-chain-compromised agent can exfiltrate keys, push to arbitrary remotes, or reach services it shouldn't.
terok-executor runs each agent in an isolated rootless Podman container with a default-deny egress firewall and a credential vault that keeps secrets on the host.
Prerequisites¶
- Python 3.12+
- Podman (rootless) —
podman machine initon macOS nft(nftables CLI)
Install and run¶
The first run interactively offers any missing prerequisites —
sandbox services, container images, agent credentials — one
[Y/n] prompt at a time. Mandatory items (services, images)
block the launch if declined; optional ones (SSH key, auth) print
the consequence and proceed.
For non-interactive environments (CI, scripts) do the bootstrap explicitly first:
terok-executor setup # install sandbox services + build base images
terok-executor auth claude # OAuth or API key
terok-executor run claude . -p "..."
setup is idempotent — safe to re-run after upgrades. Only the
image build can be run on its own: terok-executor build rebuilds
the L0+L1 images without touching the sandbox services. The other
setup steps (shield hooks, gate, credentials-DB provisioning) are
only available through setup.
Authenticate¶
terok-executor auth claude # OAuth login
terok-executor auth vibe # interactive API key prompt
terok-executor auth gh --api-key ghp_… # non-interactive
Credentials are stored on the host; containers never see real keys. See Security for the vault details.
Use as a library¶
from terok_executor import AgentRunner
runner = AgentRunner()
runner.run_headless(
agent="claude",
repo=".",
prompt="Add type hints to utils.py",
max_turns=25,
)
AgentRunner is the same entry point that terok-executor run uses
under the hood — and what
terok builds on for multi-task
orchestration.
Next steps¶
- Agents — supported agents, custom definitions, auth flows
- Security — firewall, vault, restricted mode
- Launch modes — headless, interactive, web, tool
- Vault internals — architecture deep dive