Claude Code as a chat agent: CLI models and slash commands, compact/rewind/fork/export, context and cost meter, subagents, hooks, plugins, skills, permission rules and profiles — with allow/deny for every action. Ships executable code.
Install in Lumen: add the server, then pick Claude Code from the list.
http://localhost:8730
Claude Code as a chat agent in Lumen — with the same engine as the claude command: your login, CLAUDE.md, settings and MCP servers all apply.
answers as Markdown with copyable code blocks
Bash, thefile to open; changed files open on their own
matching your version and account; ↻ next to the picker reloads them
(Haiku, for example, has none); Lumen remembers the choice per model
Plan and — only when enabled — No prompts
/ for slash commands, @ for files, ↑ for earlier messages,paste (Ctrl+V) or attach images, send the open file and selection along
The chat's "…" menu and the / palette offer, for the current conversation:
/compact [what to keep]) — summarizes the conversation to free contextfrom Claude's backups and the conversation continues as a fork from before that message (needs "Back up files for rewind", on by default)
and, with a claude.ai plan, the 5-hour and weekly limits
the output of commands such as /context
tool calls nested beneath them
hint and source (built-in, project, user, plugin, skill). /name args reaches Claude as a real slash command. /mcp lists the MCP servers and their state without sending anything.
"Always allow Bash(npm:*)" (built from Claude Code's own suggestions)
Settings → Extensions → Claude Code:
claude, default and fallback model, moremodels, model list cache, default effort, extended thinking (adaptive, budget, off), fast mode
answers, maximum steps, cost limit, additional instructions
(all, read-only, or a list), sandbox, more folders, "No prompts" mode; the extra mode "Only allowed" denies whatever no rule allows
plugins (local folders), skills, "only these MCP servers"
headers, environment variables
One per line — name | description | prompt | tools | model (tools comma-separated, model an alias, \n = line break), or JSON as for claude --agents:
reviewer | Reviews code for bugs | You are a strict code reviewer. | Read,Grep,Glob | sonnet
One per line — Event | matcher | command. Events: PreToolUse, PostToolUse, UserPromptSubmit, Stop, SubagentStop, SessionStart, Notification, PreCompact. The command gets the event as JSON on stdin and runs in the project folder; exit code 2 blocks (stderr is the reason, shown as a notice), JSON on stdout works as in Claude Code.
PreToolUse | Bash | ./scripts/check-command.sh
Quick actions (Label | prompt), commands (name | description | prompt) and profiles (Label | mode=plan | model=opus | effort=high | key=value) are filled with examples; change them freely. Prompts can use {{selection}}, {{file}}, {{clipboard}} and, in commands, $ARGUMENTS. A profile's extra key=value sets any setting of this extension for chats using it, e.g. toolset=readonly or sandbox=on. The defaults include Plan (Opus, high), Quick edit (Sonnet), Read-only review and Sandboxed autonomy; /review, /explain, /tests, /plan, /remember (adds to CLAUDE.md) and /memory. Each chat can also carry its own system prompt.
New project → Claude Code project creates CLAUDE.md, .claude/settings.json, the slash commands /review and /explain and a reviewer subagent.
Claude Code must be installed and signed in:
npm install -g @anthropic-ai/claude-code
claude
The extension ships executable code that must be confirmed on installation. It starts the installed claude.
claudePath — Path to claudemodel — Default modelfallbackModel — Fallback modelshowLegacyModels — Offer older modelsextraModels — More modelsmodelCacheMinutes — Cache model list (minutes)effort — Default effortthinking — Extended thinkingthinkingBudget — Thinking budget (tokens)fastMode — Fast modestreaming — Stream answers livemaxTurns — Maximum stepsmaxBudget — Cost limit per message (USD)appendSystemPrompt — Additional instructionsoutputStyle — Output stylepersistSessions — Save sessionsfileCheckpoints — Back up files for rewindallowedTools — Tools allowed without askingdisallowedTools — Denied toolsaskTools — Tools that always asktoolset — Built-in toolscustomTools — Tool listsandbox — SandboxadditionalDirectories — More foldersallowBypass — Allow "No prompts" modecustomAgents — SubagentspluginDirs — Pluginsskills — SkillsstrictMcp — Only the MCP servers listed herehooks — HooksuserSettings — Load user settings (~/.claude/settings.json)projectSettings — Load project settings and CLAUDE.mdlocalSettings — Load local settings (.claude/settings.local.json)mcpServers — Additional MCP servers (JSON)betas — Beta featuresenv — Environment variablesquickActions — Quick actionscustomCommands — Commandsprofiles — Profiles