Tools and permissions
Files, web, browser, shell, documents, voice, sub-agents and MCP, and the rules that decide what runs without asking.
Built-in tools
| Tool | What it does | Default |
|---|---|---|
read_file, write_file, patch, list_dir, search_files | Files. Documents, spreadsheets and PDFs come back as text | on |
extract_archive | Unpack zip, rar, 7z, tar, gz safely | on |
create_document | Make .docx, .xlsx and .csv files | on |
send_file | Send a file to you in the chat (up to 50 MB) | when a channel runs |
web_search, web_extract, http_get | Search, readable pages, raw APIs | on |
todo | Task list for multi-step work | on |
delegate_task | Up to 6 sub-agents in parallel | on |
skills_list, skill_view, skill_manage | Find, read and write skills | on |
memory, session_search | Long-term memory, search past conversations | on |
cronjob | Schedule agent runs and script watchers | on |
clarify | Ask you a question with choices | on |
send_message | Message you on a chat channel | when a channel runs |
vision_analyze | Understand images and screenshots | on |
transcribe_audio | Speech in audio or video files to text | when [stt] is set |
terminal, execute_code, process | Shell and Python, background processes | [tools] terminal = true, asks first |
browser_* | A real Chromium browser with saved logins | [browser] enabled = true |
image_generate | Create images | when [image] is set |
text_to_speech | Audio from text | when [tts] is set |
mcp_<server>_<tool> | Every tool of every MCP server you add | per server |
Permission rules
| Setting | Effect |
|---|---|
files = "workspace" (default) | File tools stay inside the workspace folder |
files = "full" | Reads anywhere; writes outside the workspace and reading secret files (.env, ~/.ssh, *.pem, …) ask first |
terminal = false (default) | No shell, no code execution |
terminal = true, confirm = true | Every command asks first (Telegram buttons, or yes/no) |
confirm = false | Commands run without asking. Only on a machine you trust |
[mcp] confirm = true | Every MCP tool call asks first |
Files it can read
PDF, Word, Excel, PowerPoint, OpenDocument, CSV, images, and archives (zip, rar, 7z, tar). Spreadsheets come back per sheet; scanned PDFs become page images for the vision model. Old binary .doc/.xls need LibreOffice; RAR and 7z need unrar, unar or 7-Zip (the Docker image has them). smooth-agent doctor shows what your machine can read.
MCP servers
bash
smooth-agent mcp add github --env 'GITHUB_PERSONAL_ACCESS_TOKEN=${GITHUB_TOKEN}' -- npx -y @modelcontextprotocol/server-github
smooth-agent mcp add docs --url https://example.com/mcp --header 'Authorization=Bearer ${DOCS_TOKEN}'
smooth-agent mcp list
smooth-agent mcp tools githubLocal (stdio) and remote (Streamable HTTP / SSE) servers both work. ${VAR} is read from your environment, so tokens never sit in mcp.json.
Last updated: