Install the Claude Code integration to see Claude's state in the sidebar and to answer its requests from the inbox, from another pane, or from your phone.
What you get
| Feature | How it works |
|---|---|
| State | Claude Code hooks report when a turn starts, when a tool runs, when the agent is idle, and when it fails or ends. The sidebar shows this state. |
| Permission requests | A permission request becomes an inbox card. Vibeke sends your answer back through the hook, so Claude continues without a keystroke. |
| Plan reviews | A plan review becomes a card. You answer it through the hook. |
| Questions | A question from Claude becomes a card. Vibeke answers it by sending keystrokes to the Claude prompt. This is best effort. |
| MCP input requests | The installer also adds the hook for an MCP server that asks you for input. |
| Background tasks | Claude reports its background shells and monitors. The sidebar shows them after ⧗. See Agents and interactions. |
| Usage and limits | Vibeke reads Claude's transcript file. It records tokens and cost for each turn and notes rate limits. Run vibeke agent turn-usage <run> or vibeke agent limits. |
| Resume | Vibeke starts Claude with a session ID. After a restart it can resume the session with claude --resume. |
Claude Code has no structured model switch in Vibeke. vibeke agent models reports that it is not supported. Type /model in the pane instead.
Install
- Open Setup in the command palette. Select Claude Code. Review the planned change. Confirm.
- Or run the commands below. Without
--yes, the command only prints the planned change.
vibeke integration install claude
vibeke integration install claude --yes
- Start
claudein a Vibeke pane. A session that already runs needs a restart to load new hooks.
Files it changes
~/.claude/settings.json, orsettings.jsonin the directory named byCLAUDE_CONFIG_DIR.- Vibeke adds one hook entry for each event it watches. Each entry runs
vibeke hook claude <Event>. A_vibekekey marks the entries as its own. - The installer keeps your other hooks and your other settings.
- Before the first change, it saves a copy named
settings.json.vibeke-bak-<timestamp>next to the file. ~/.claude/skills/vibeke/SKILL.mdholds the Vibeke skill. A skill is a file of instructions that Claude reads when a task needs it. This skill teaches Claude thevibekecommands for panes, agents, tasks, previews, the agent browser and images. A line after the header of the skill marks it as managed by Vibeke. A reinstall overwrites it. The installer does not replace avibekeskill without that line.
The --mcp flag registers the vibeke mcp server in ~/.claude.json. It writes only that entry. The server offers previews, the agent browser and images as MCP tools.
Check
vibeke integration status claude
vibeke integration doctor claude
vibeke integration capabilities claude
statussaysinstalled,partialornot installed. For a partial install it lists the missing events.doctorfinds theclaudeprogram, prints its version, and compares it with the validated range. It also shows the capabilities that apply.capabilitiesprints the same capability list for your version.- These checks are local. To confirm that hooks reach the server, start
claudein a Vibeke pane and runvibeke agent list. The run should show"source": "Structured"underexecution. The valueScreenmeans that Vibeke reads only the terminal output.
A Claude Code version outside the validated range gets observe and keystroke answers only. Vibeke still shows state. It does not send native answers until the version is validated.
Settings
In config.toml:
[agents.approvals.claude]
persist_always = false # true: "allow always" edits Claude's settings
persist_destination = "localSettings"
By default, an "allow always" answer applies only to the session. With persist_always = true, Vibeke writes the rule to the Claude settings file that persist_destination names. See the configuration reference.
Remove
vibeke integration uninstall claude --yes
This removes only the entries that Vibeke added. Your own hooks stay.
Limits
- Hooks only report for Claude Code that runs in a Vibeke pane.
- A hook waits up to 30 minutes for your answer. After that, Vibeke gives no decision and Claude shows its own dialog.
- Vibeke answers Claude questions with keystrokes, not through a hook.
- Vibeke does not change Claude's model. Use
/model.
Next steps
- Agents and interactions: how requests, answers and background work behave for every agent.
- Run agents in parallel: start several Claude Code agents and track them.
- Troubleshooting: fix hooks that do not report.