200agent gives you a team of Claude agents on your Mac. A lead splits your goal into cards, named agents work in parallel in live terminals, a second model checks the result, and nothing risky happens without your approval.
How it works
You write what you want in plain words. The lead plans it and asks you once. The crew works in parallel, a second model checks the result against the acceptance criteria, and you get a report with evidence. Every file change can be undone.
Risky actions such as deleting, sending or paying always wait for your yes.
+ Total sales: 184,500 TL (+12%)
+ Best seller: Ankara branch
✓ numbers recomputed
✓ reviewer: passed
→ raporlar/ANA-1-ada.md
Real flow from the app, using sample files.
Demo
A screen recording of the app, 1.5× speed, on sample files: a goal typed to the command bar, the lead's plan card, three agents working at once, and a delete that stops for approval.
Recorded with an offline demo model so no real accounts or costs are involved; the app, office, board and terminals are the real product.
What's built today
Everything on this board runs in the current build and is covered by automated tests, including real process-kill recovery tests.
Plans arrive as cards you approve; a kanban board stays in step with every checkpoint.
17 roles, team templates, per-agent models and a direct line to each agent while it works.
Checkpoints, an outbox for side effects and a watchdog: finished steps never rerun.
Web, e-mail and file text is marked untrusted; injection patterns are flagged and risky calls ask first.
A learning router picks the cheapest model that clears a measured quality bar, and escalates when needed.
Word, Excel, PowerPoint, PDF, a real browser with click approvals, and MCP for GitHub, Sentry, Supabase, Stripe and PostHog.
Built on Claude
200agent assigns Claude models by role and measures how well each one does. Prompts are laid out for prompt caching, and every call is logged with tokens, cache hits and the reason it was routed. API credits would fund the evaluation runs that tune the router and the public edition on the Claude API.
Turns a vague goal into a plan with acceptance criteria and step difficulty.
Tool use in a bounded loop: files, shell, documents, browser.
Checks the work without seeing how it was produced; another model family when one is connected.
Settles disagreements on high-risk steps such as payments or important e-mail.
Quick answers, classification and summaries for context compaction.
Web200 is a web design and development studio in Istanbul that builds its own products: 200news for news publishers, and now 200agent. 200agent is an agent office that a non-programmer can run, that checks its own work, and that never spends money or touches anything important without asking first.
Founder: Umut Boyludag · web200.com.tr · Contact