Work Mode Configuration

Every Work Mode capability lives in one control panel — hover the robot icon in the input box toolbar to open it. This page walks through each setting: what it is, where to configure it, its default, and whether it's stored globally or per-session.

The Control Panel at a Glance

The panel has three sections top to bottom: Mode, Built-in (Web Search, Memory, Code Execution), Extensions (Skills, MCP, Knowledge Base, Working Directory). Hover a row under Built-in / Extensions and a sub-panel slides out to the right for further settings. Code Execution, Skills, MCP, and Working Directory only work in Work Mode (dimmed in Chat Mode); Web Search, Memory, and Knowledge Base work in both modes.

Per-session vs Global: per-session settings only affect the current chat and reset to default in a new one; global settings (Skills and MCP enablement) are configured once and apply everywhere.

Mode

The top of the panel has two side-by-side buttons — Chat Mode (chat bubble icon) and Work Mode (briefcase icon). Click one to use it. The selected one is filled brand blue; the other is outlined. A one-line description of the selected mode sits underneath.

• Chat Mode — no tools injected. Best for quick Q&A, writing, translation, and explanations.
• Work Mode — the full tool set. Best for multi-step tasks with files, code execution, tools, MCP, skills, or knowledge bases.

Smart Switching

The toggle below the mode buttons shows while Chat Mode is selected, and only takes effect before you send the first message. Its scope is a global default — you're really setting "whether new chats start with Smart Switching on."

• On (default) — when you send the first message, the AI judges whether the task suits Work Mode better; if so, it drops a "Work Mode suggested" card into the message flow for you to accept with one click.
• Off — always stick to the mode you picked, no interruptions.

After the first message, the toggle shows "Only available before the first message" and dims — Smart Switching only makes sense at the start of a chat.

Built-in Capabilities

Web Search

Lets the AI search the web and read page content. The row has a switch on the right; when on, the current provider name shows next to the label. Hover the row and the sub-panel lets you switch providers.

• Reading page content: only Chatbox AI and Tavily support it — they add a parse_link tool so the AI can open a link and read the full text. Bing does search only.
• Scope: the Web Search toggle is per-session. New chats on a Chatbox AI model default to on; other models default to off.
• Relation to Chat Mode: Web Search works in both modes from this same row. The gear in the sub-panel header jumps to Settings → Web Search to enter API keys or your SearXNG instance URL.

Memory

Lets the AI keep durable facts it learns as you work together and use them in later chats. The badge on the right shows which memory is in use — Global Memory, Copilot Memory (shared by every chat with the same Copilot), or Off. The gear in the sub-panel header opens memory settings, where you can review and manage saved memories.

Code Execution

The core of Work Mode. The AI can run Node.js or Bash in a local sandbox, with a full set of read / write / edit / search-file and export-download tools. The Badge on the right shows the current approval level; the sub-panel switches it.

• Scope: the approval level is per-session.
• Sandbox boundary: code runs in a local sandbox directory, isolated from your real filesystem. Relative paths inside the sandbox working dir are written directly; writing an absolute real-filesystem path needs approval (unless Full Access is on, or that path is in a Working Directory below).

When a command needs approval, its step turns amber in the timeline and the input box is replaced by an approval bar showing the exact command. Approve or deny it there; "View" jumps to the step.

Extensions

Skills

Skills are "expert playbooks" that tell the AI how to do a class of tasks better (based on the agentskills.io spec, compatible with Claude Code skills). Once enabled, the AI loads the full instructions on demand via load_skill. The Badge on the right shows how many are enabled.

• Where to manage: Settings → Skills. Install from the skills marketplace, install from a GitHub repo, or open the local skills folder to drop one in manually.
• Compatible sources: besides what you install, Chatbox auto-discovers skills in Claude Code's ~/.claude/skills and the shared ~/.agents/skills directory.
• Scope: skill enablement is global (enabledSkillNames), applies to all chats, with no limit.
• Structure: each skill is a folder with a SKILL.md whose frontmatter has name (lowercase + hyphens), description, etc.
• Trigger in chat: typing / in the input box suggests skill commands; you can also click a skill in the panel list to insert /skillName into the input.

MCP

MCP (Model Context Protocol) connects the AI to external tools and services. Chatbox ships a set of cloud MCP servers and also lets you add your own. Tools are injected per server. The Badge shows the count of enabled servers; toggle each in the sub-panel.

Built-in cloud MCP (requires Premium):

Custom MCP: add it in Settings → MCP. Two transport types:

• Remote (http/sse) — enter a service URL; works cross-platform (desktop / Web / mobile).
• Local (stdio) — enter a local command, args, and env vars; desktop only (it launches a process on your machine).

Knowledge Base

Bind a knowledge base and the AI can do semantic retrieval over it and read documents by passage, answering from your own material (RAG). The row shows the bound knowledge base name; the sub-panel switches it (single choice).

• Create it: Settings → Knowledge Base. On creation you configure:
  • Embedding model (required) — turns text into vectors; add it first under Settings → Provider → Model List.
  • Rerank model (optional) — makes retrieval more accurate.
  • Vision model (optional) — preprocesses image files; needs a vision-capable model.
  • Document parser (optional) — processes uploaded documents; can use MinerU, etc.
• Supported formats: PDF, DOCX, TXT, MD, RTF, PPTX, XLSX, CSV, EPUB, and more. Added files are vectorized automatically.
• Binding: a chat binds one knowledge base, per-session. Once bound the AI gets query_knowledge_base, read-by-passage, and related tools.
• Platform limits: Windows ARM64 doesn't support the knowledge base (libsql limitation); on Web / mobile, local file parsing is limited to text formats.

Working Directory

Grant the AI a real local folder and it can read and write in it directly, without per-action approval — like a pair programmer working inside your project. The Badge shows the folder count; add/remove in the sub-panel.

• Platform limits: available on desktop only.
• No-approval semantics: paths inside a working directory behave like the sandbox — reads and writes go through directly, no approval card. Absolute writes outside the directory still need approval (unless Full Access is on).
• Scope: per-session. For a new chat (before the first message) it's staged, then written into that chat's settings on first send.
• The "Add Folder" button at the bottom opens the system directory picker.

Step-limit pause

To keep the AI from drifting on long tasks, it pauses after 25 consecutive tool calls by default so you can confirm direction, then "Continue" or "Stop" from the bar that replaces the input box.

• Turning it off: the arrow next to "Continue" offers "Continue, and don't pause this chat again" and "Continue, and don't pause any chat again".
• Where it lives: the global default is Settings → Chat → Pause after every 25 steps; each chat can override it in its conversation settings.
• Full Access: chats at Full Access don't pause by default, unless the chat's own setting turns it back on.

Quick Reference