Describe the job in plain language. The agent plans it, runs it on your machine, and delivers the result — not just advice.
Chat · Do modes
6 LLM providers
Safe local execution
Files · Office · Images
You ask. It does the work.
Chatbots stop at suggestions. AI Work Agent executes — finding files, transforming documents and images, producing outputs under a sandboxed attempt folder, and retrying when a step fails.
Talk or do
Each turn chooses chat for Q&A or do-mode to carry out the task on your PC. Force either with /chat or /script.
Runs on your machine
Work happens locally in the project virtualenv. Reads can go anywhere; writes stay inside the current attempt folder so originals stay read-only.
Ready-made helpers
Built-in tools for Excel, Word, PowerPoint, PDF, and images are auto-cataloged so common jobs finish without ad-hoc package hunting.
Execution loop
Plan, run, repair — until the job is done.
The agent streams its plan, carries out the steps, installs missing packages after confirmation when needed, and can automatically repair and retry up to a configured limit.
Streaming plan with a Thinking indicator
Optional confirm before packages or file-changing work
Automatic repair retries on failure
Safe execution
It works on copies — originals stay untouched.
Before run, static checks flag delete, write, and rename risks. At runtime, filesystem guards confine changes to the attempt directory. Transforms read sources and save new artifacts — Downloads and Desktop files are never overwritten in place.
Pre-run risk scan for destructive actions
Write root under output/<timestamp>_attemptN/
Optional consent when file changes are detected
What it can do
Files, Office docs, and images — end to end.
Attach folders with /attach, then ask for real outcomes: resize photos, reshape spreadsheets, extract PDF text, build slides. Helpers under agent_tools/ register themselves into the agent catalog at runtime.
Excel · Word · PowerPoint · PDF · Image helpers
/attach paths injected into context
Persistent memory.txt for standing preferences
How it works
1
Describe the job
Say what you need, optionally /attach files or folders, and pick a model with /model.
2
Agent plans the work
It chooses chat or do-mode, streams the plan, and uses conversation history for follow-ups like “do the same for the next folder.”
3
It runs and delivers
Work executes locally; results land in the attempt folder. On failure it repairs and retries within the limit you set.
A local agent that finishes the task
OpenAI, Claude, Grok, Cerebras, Kimi, and DeepSeek via .env. Windows, macOS, and Linux with setup/run scripts that keep everything inside .venv. Python is the engine under the hood — the product is getting the work done.
Development / My Role
I design and build AI Work Agent end to end — the ask-and-execute loop, streaming CLI, multi-provider LLM client, local execution sandbox and safety scans, auto-discovered tool registry, and the operator workflow around history, attachments, and saved runs.
Built with Python, OpenAI & Anthropic SDKs, prompt_toolkit, pandas/openpyxl, python-docx, python-pptx, pypdf, and Pillow.
