Assistant capabilities

GoWork's assistant is a resident, LLM-driven agent that runs on your machine. It doesn't just chat — it reasons about a task, calls real tools, reads the results, and keeps going until the work is done or it needs your confirmation. Here is what it can actually do today.

Runs tasks, step by step

The assistant works in a reason–act loop: it thinks, calls a tool, reads the result, and repeats. A single run can chain dozens of tool calls to finish a multi-step job. Every run has a clear outcome — done, waiting for you, waiting on a delegated runtime, or stopped — and can be cancelled mid-flight. Long-running and desktop tasks queue safely, so two jobs never fight over the same resource.

Acts directly with real tools

Inside a guarded workspace, the assistant can:

  • Files — read, write, and edit files; list, glob-search, and grep across a directory.
  • Shell — run commands and capture their output.
  • Web — search the web and fetch pages, with several search engines as fallbacks.
  • Documents — read PDFs and Office files (Word / PowerPoint / Excel) and plain text.
  • Images — generate images from a prompt, and view local images as input.

File and shell access is confined to the workspace; read access to other folders is granted explicitly and remembered for the session.

Delegates the coding to runtimes

For real programming work, the assistant hands the task to a coding runtime — Codex or Claude Code — and manages the session: streaming progress, injecting follow-up input, continuing, or cancelling. It then rewrites the runtime's raw output into a plain-language answer for you.

Remembers across sessions

The assistant keeps long-term memory at four scopes — you (global preferences), project, conversation, and run — and infers where a fact belongs from how you phrase it ("from now on…", "for this project…"). It can recall past tasks, resolve references like "that thing from last week", and promote a useful how-to into a reusable Skill.

Schedules and automates

Give it a recurring job in plain language and it creates a scheduled task — once, daily, weekly, monthly, or yearly. At the trigger time it either sends you a reminder or wakes itself to carry the task out. A local scheduler recomputes the next run correctly across month-ends and daylight-saving changes, and recovers tasks after a restart.

Reaches you over IM

Pair a channel and the assistant comes to you: DingTalk, Feishu, and Telegram are fully supported, and it can push text and images back to the conversation. It can bind a chat to a running task so you can keep a job going from your phone. (WeChat is an experimental integration and currently handles text and transcribed voice only.) See Channels.

Extends through Skills and MCP

  • Skills — reusable playbooks the assistant discovers and activates by topic, or that you invoke explicitly.
  • MCP — connect Model Context Protocol servers, and the assistant can list and call their tools and read their resources directly, over stdio or HTTP.

Automates the desktop (optional)

When you enable desktop control, the assistant can drive GUI apps: find windows and controls, click, type, use hotkeys, scroll, capture a window, and wait for a control, window, file, or process before acting. Backends cover Windows, Linux, and macOS.

Stays safe, asks when unsure

  • Tool policy — a default-deny gateway checks every tool call; read-only actions pass, risky ones require confirmation, and unknown tools are refused.
  • Approval governance — when a delegated runtime asks for permission, saved policies auto-approve or auto-deny; otherwise it asks you, and can remember "always allow" for this session.
  • Clarification — questions, confirmations, and pending answers are first-class, so the assistant asks rather than guesses.
  • Resilience — the models powering the assistant sit behind a circuit breaker with automatic failover and recovery.

Multimodal input

Beyond text, the assistant accepts voice (speech-to-text, with OpenAI or Gemini as the backend), reads documents, and both generates and views images.


See also: Model gateway · Channels · Reference

Ready to try it? Download GoWork or request access.