Extensions & MCP

Four systems decide what the agent can do, under which persona, and which external tools it can reach.

Agents (prompt presets)

Entry: /agents. An agent = system prompt + model preference + tool allowlist + per-tool approval policy.

Item	Details
Categories and search	Group by purpose (coding, writing, analysis …) and search by name
Tool allowlist	Restrict which tools the agent may call
Approval policy	Per tool: silent / auto / ask
Import / export	JSON import and export of all agents for backup or sharing

Skills

Entry: /skills. A skill is a reusable prompt template, invoked in chat with /skill-name.

Source	Details
Built-in	Translate, summarize, explain, polish — available immediately
Custom	Your own prompt template and system prompt; test output right on the page
On-disk directories	Load Markdown / JSON skill files from configured directories — good for sharing via a repository

Tools

Entry: /tools. Built-in tools can be toggled; disabled tools are not callable.

Tool	Purpose
Note search / read / create / update	Read and write notes from chat, with conditional search
Web search	Fetch external information
Memory search / create	Query and write long-term memory
Graph search	Query entities and relations
Todo list	Emit a structured plan and check items off live
Command execution	Run commands in the project workspace (permission-mode governed)
File operations	List, read, write, move, delete
Ask question	Interactive question when a decision is needed

MCP servers

Entry: Settings → MCP server. Connect external Model Context Protocol services to bring their tools into chat.

Item	Details
Transport	`stdio`: spawns a child process and speaks JSON-RPC 2.0 (fully supported). `sse`: configurable, not wired up yet
Connection config	JSON — `command` / `args` / `env` for stdio, or `url` for sse
Tool discovery	Calls `tools/list` and merges discovered tools into the chat toolbelt
Tool invocation	Calls `tools/call`; arguments and results render structurally in the message stream
Enable / disable	Per server, with ordering

MCP tools share the same permission modes and approval policies as built-in tools; run a new server in “ask” mode first if you are unsure what it does.

Configuration examples

MCP server (stdio)

{
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "E:\\repos"],
  "env": { "NODE_ENV": "production" }
}

Skill file (on-disk directory)

---
name: code-review
description: Review changes against a fixed checklist
---
Review the selected code in this order:
1. Correctness and edge cases
2. Resource cleanup and concurrency
3. Naming and readability
4. Consistency with the surrounding module

Configure the skill directories in settings; a file with the same name as a database skill wins.

Agent (import / export JSON)

{
  "name": "Code reviewer",
  "category": "coding",
  "systemPrompt": "Report only real defects, each with a minimal reproduction and a fix.",
  "allowedTools": ["read_file", "search_code", "run_command"],
  "toolPolicies": { "run_command": "ask", "update_note": "silent" }
}

Apps (embedded browser)

Entry: /apps. Register frequently used pages as “apps” and open them inside the app instead of switching windows.

Item	Details
Registration	Name plus URL (a local service such as `http://localhost:5174`, or any site)
Opening	Embedded view with address bar, back / forward, reload and quick links
Timeout hint	When a site refuses embedding via `X-Frame-Options` / CSP, a hint appears after 4 seconds suggesting the system browser