VibePod CLI 0.25: Cline, Junie, Cursor and Warp

2026-11-13 · Release · CLI · Agents

Four new agents

vp run now starts nineteen agents. Each of the four new ones runs from an image that pins one release and switches off the agent's self-update, and keeps its state under ~/.config/vibepod/agents/<agent>/, so a login outlives the container.

Cline

vp run cline, or vp cl, starts Cline CLI, the terminal edition of the Cline coding agent. It shares its agent core with the VS Code extension: plan and act modes, MCP servers, rules, skills and checkpoints. vp run cline -- auth opens the setup for a Cline account, a ChatGPT subscription or your own API key, and the choice is kept in the config mount. Tasks always run in act mode, because a task left in plan mode would never edit a file.

Junie

vp run junie, or vp ju, starts Junie, the LLM-agnostic coding agent from JetBrains. Log in with your JetBrains account from the TUI, or set JUNIE_API_KEY, which also works for tasks, where there is no TUI. With your own keys, Junie talks to Anthropic, OpenAI, Google, Grok, OpenRouter or a LiteLLM proxy. Junie's bundled Java runtime ignores SSL_CERT_FILE, so the image imports the proxy's CA into the runtime's trust store and Junie's HTTPS calls pass through vibepod-proxy.

Cursor

vp run cursor, or vp cu, starts Cursor CLI, the terminal agent from the makers of the Cursor editor. Models come through your Cursor account. Log in once with vp run cursor -e NO_OPEN_BROWSER=1 -- login and open the printed URL in the browser on your host, or set CURSOR_API_KEY. A detached task cannot answer Cursor's workspace-trust prompt, so task mode marks the workspace as trusted; command approvals stay in place unless you add --ikwid. Cursor talks to its backend over long-lived HTTP/2 streams, which vibepod-proxy 0.6.0 passes through instead of buffering, so this release requires that proxy version.

Warp

vp run warp, or vp w, starts the Warp Agent CLI, which runs the agent from the Warp terminal app in any terminal. On the first start it shows a verification link and a code; open the link in the browser on your host, or set WARP_API_KEY to skip the login. Warp keeps its login only in a system keyring, which a container does not have, so the image starts a private one with its files on the config mount, and the login is remembered across runs. Warp is interactive only: it takes no prompt argument and has no ACP server, so vp task create warp and --acp are not available.

Tasks, editors and skills

Cline, Junie and Cursor also run as background tasks, and inside your editor with --acp, which now covers fourteen agents:

vp task create cline "fix the failing test" --ikwid
vp task create junie "summarize the README"
vp task create cursor "summarize the README"

Skills installed with vp skills are mounted for all four. Each of them brings its own model setup, so VibePod's llm settings and --provider do not apply to them.

Bring the setup you already have

vp import copies an agent's configuration from your home directory into a VibePod profile, mapped to the paths the agent reads inside its container: settings, custom models, MCP servers, hooks, skills and memory files. Credentials and session history stay behind unless you opt in:

vp import                      # scan your home directory for installed agents
vp import claude --dry-run     # resolve the plan, write nothing
vp import claude --to-profile work --create-profile
vp import claude --with-credentials

With --from-profile, the same command copies one profile's agent configuration into another.

Mounts at host paths

A single absolute path is now enough to mount a directory at the same place in the container: vp run claude -v /srv/datasets. For tools and configs that refer to absolute project paths, --workspace-system-path mounts the workspace at its host path as well as at /workspace. On SELinux-enforcing hosts such as Fedora, VibePod prints a hint when its bind mounts keep a label that containers may not access, and selinux_relabel: true opts in to relabeling the mounts VibePod owns. Relabeling is permanent, so read the SELinux notes first.

Smaller additions

vp skills export copies installed skills out as plain SKILL.md folders, so you can commit them to a repository or hand them to someone else.
A provider with an API key can use a plain http:// endpoint after an explicit opt-in, for LAN servers such as llama-server --api-key. The opt-in stays in your local configuration and is not part of a provider export.
Session and task records keep the image tag, the image ID and the agent version, so you can tell later which build did a run.
Python 3.15 is tested and declared, and the docs cover uv tool install vibepod.
The docs have new pages on external tools (LiteLLM, Headroom, RTK, Graphify) and on adding an agent, and describe how to keep Claude's project memory in the repository.

Fixes

macOS: an agent started inside a herdr pane no longer fails to launch. The herdr socket cannot be bind-mounted off Linux, and VibePod now skips that mount on every engine there, Docker Desktop and Colima as well as Podman machine.
Rootless Podman: Hermes starts. Earlier releases rejected the launch; VibePod now maps your host user onto Hermes's runtime user, so the files in your workspace stay yours. This needs Podman 4.3 or newer.
Native Windows: attaching to an agent and --acp work over the named pipes of Docker Desktop and Podman Desktop, where they crashed before, and keystrokes are no longer doubled. A few keys do not reach the agent yet; see the Windows notes.
Pi tasks: vp task create pi works. Pi runs in its JSON mode, and tasks trust the project's local files unless you have chosen a trust setting yourself.

In testing: board automation

Board automation and its review workers ship with 0.25.0, but they are still in testing, and commands and options may still change.

vp board work connects VibePod to a vibepod-board project. It claims the next planned task whose dependencies are done, checks the task's branch out in a clone next to your repository, and runs the agent headless in that clone with the task as its prompt. An optional verify command, such as the test suite, has to pass before the task moves to Review:

vp config allow-dir ~/src/app       # once per repository
vp board work VP --agent claude \
  --repo ~/src/app \
  --verify "python -m pytest -q" \
  --ikwid

Every run adds a report to its task: the agent's summary, its commits, the verify output and the duration. An agent that finds a task unclear asks instead of guessing: the task is blocked with the question on its card, and the next run gets the answer in its prompt. With --poll 2m the worker waits for new tasks instead of exiting. The board token is never passed to the agent's container.

Review workers and the pull request

Started with --mode review, a worker takes the tasks in Review instead and decides whether each one needs rework or is ready for its pull request. Several reviewers, with different agents, can review the same task. Feedback reaches the implementing agent in its next prompt, and the rework continues on the same branch:

vp board work VP --agent codex --repo ~/src/app --poll 2m
vp board work VP --agent claude --repo ~/src/app --poll 2m --mode review

Workers commit on local branches. Once a task has the approvals it needs, the approving worker pushes the commit it approved to the task's branch on GitHub, and Open PR on the board is prefilled with the title and description the implementing agent suggested. A push is never forced, uses the git credentials of the user the worker runs as, and can be turned off with --no-push.

Upgrading

python -m pip install --upgrade vibepod
docker pull vibepod/proxy:latest

Proxied runs need vibepod-proxy 0.6.0 or newer. VibePod pulls the proxy image only when it is missing, so pull it once if you already have an older one, with docker or podman. If a proxy is still running, stop it with vp proxy stop first. Mistral's agent is now called vibe; vp run devstral, existing agents.devstral configuration and VP_IMAGE_DEVSTRAL keep working. Session and task databases are migrated on first use.

Release details

New agents cline, junie, cursor and warp.
New vp import for host agent setups and profile-to-profile copies.
Same-path volume shorthand, vp run --workspace-system-path and opt-in SELinux relabeling.
New vp skills export copies installed skills out as plain SKILL.md folders.
A provider with an API key can use a plain http:// endpoint after an explicit opt-in.
Session and task records keep the image tag, the image ID and the agent version.
The herdr socket mount is skipped on every container engine off Linux.
Hermes runs on rootless Podman.
On native Windows, attaching to an agent and --acp work over named pipes.
vp task create pi works: Pi runs in its JSON mode.
vibepod-proxy 0.6.0 or newer is required. A custom proxy image needs the label io.vibepod.proxy.version.
Python 3.15 is tested and declared, and the docs cover uv tool install vibepod.
New docs pages: external tools (LiteLLM, Headroom, RTK, Graphify) and adding an agent.
In testing: vp board work with implementing and reviewing workers, configured under board: or with VP_BOARD_URL and VP_BOARD_TOKEN.
Thanks to @janpipek (SELinux relabeling), @ReimarBauer (Windows attach), @sguter90 (profile docs) and @Qrytics (proxy policy validation) for their contributions.