Give it a goal. Watch it work. Get back results.
Finisher is an autonomous AI agent that takes a plain-English goal and executes it end-to-end — searching the web, writing and running code, reading and writing files — then delivers the finished result. Every step is visible in real time.
You type a goal. Finisher plans it, executes it in steps, and shows you exactly what it's doing the whole time.
You: "Research the top 5 AI coding tools and summarize their pricing"
Finisher: → Plans 4 steps
→ Searches web for each tool
→ Extracts pricing from each site
→ Writes and runs summarization code
→ Delivers a formatted report
No black box. Every action logged. Every result shown. Sessions are replayable step by step.
- Search the web — Tavily-powered research on any topic
- Write and run code — generates Python, executes it in a sandbox, uses the output
- Read and write files — manages files in a sandboxed workspace
- Automate the browser — Playwright integration for scraping and navigation
Gradio split-screen: your goal on the left, the agent's live execution on the right. Watch it think, plan, and act in real time. Come back later and replay any session from the beginning.
Prerequisites: Python 3.11+, Anthropic API key. Tavily API key optional (enables web search).
# Install
make install
# Configure
cp .env.example .env
# Add ANTHROPIC_API_KEY and TAVILY_API_KEY to .env
# Run
python launch_ui.py
# Opens at http://localhost:7860Your Goal
↓
Orchestrator
↓
Agent Loop: Analyze → Plan → Execute → Observe → repeat
↓
Tools: web search · code execution · file ops · browser
↓
Event Stream (session.jsonl) — full replay log
↓
Finished Result
CodeAct architecture — the agent generates Python code as its primary action mechanism, executes it, observes the result, and decides next steps. Actions are auditable because they're just code.
Python · Anthropic Claude · Gradio · Playwright · Tavily · CodeAct
Built in November 2025 to understand autonomous agent architecture from the inside — before "autonomous agents" had a mainstream name. The core bet: an agent that works end-to-end and shows its work is more useful than one that asks clarifying questions at every step.
Same idea as Manus, OpenAI Operator, and Claude Computer Use. Built as a personal project to learn the architecture by building it.
Chris Park — Senior PM, Microsoft Office of the CTO, AI Incubation group. Engineering degree from Waterloo. 17 years shipping product.