VisionClaw Agent
Open-Source Multi-Tenant AI Agent Workspace — Documents, Research & Workflows
Built for agencies, operators, and founders who want an always-on AI operations team they own and host themselves.
Created by Robert Washburn | huskyauto@gmail.com | Live demo: agenticcorporation.net
Not affiliated with the unrelated AR/wearable "VisionClaw" project. This is the AI agent platform.
⚠️ Autonomous Pipelines Disclaimer (please read before forking)
VisionClaw ships several autonomous decision pipelines — the multi-model jury (jury_triage), the Agentic CI Self-Healer, and downstream implementer hooks. In this public release these pipelines are safe-by-default:
- Auto-apply is OFF. The jury votes and writes a human-readable decision log for your review, but it does not automatically queue fixes for code-mutation or apply verdicts to your codebase.
- To opt in, set
JURY_AUTOAPPLY=1 in your environment. This enables the implementer-pickup seam (data/jury-decisions/queue.json). Only enable this if you have read the code, understand the risk model, and have your own test/rollback discipline in place — autonomous code-mutation can corrupt data, break auth, leak secrets, or commit broken code if your guardrails aren't set up.
- No warranty. This project is provided AS-IS under its open-source license. The maintainers, contributors, and any platform hosting the source (including Replit) are not responsible for outcomes when you enable autonomous pipelines in your own deployment. The defaults are deliberately conservative; if you change them, you own the consequences.
- Reporting: if you find a safety bug in the gated-off default path (i.e. something autonomous happens without
JURY_AUTOAPPLY=1), please open a GitHub issue or email huskyauto@gmail.com.
What Is This?
VisionClaw Agent is an open-source, multi-tenant AI platform where 16 specialized agents work together to produce real deliverables — research reports, legal documents, financial models, marketing campaigns, slide decks, spreadsheets, and PDFs.
Instead of a single chatbot, you get a full agent workforce. Give it a task. The right agent picks it up, selects the right tools, coordinates with other agents when needed, and delivers a finished result. Every decision is traceable, every action is governed, and every integration degrades gracefully when not configured.
Fork it. Configure your API keys. Deploy. You have an AI operations team.
The app runs with just one LLM key and a Postgres database. Everything else — email, payments, voice, Drive — is optional and appears automatically when you add the key.
Roughly 180k lines of TypeScript across 460+ files. 40+ pages. 391 tools · 126 active capabilities · 16 personas · 195 tables · 41 governance rules · 573 production indexes · 6 AI providers · 6 deployment targets. Live, always-current counts: docs/CURRENT_PLATFORM_TOTALS.md. Browsable indexes: docs/tools.md · docs/personas.md.
Who is this for?
- Agencies — give every client a multi-agent operations team without burning headcount.
- Operators & founders — replace 5-10 SaaS subscriptions with one agent workforce you own.
- Solo entrepreneurs — run a research / content / sales / finance team of one.
- Research teams — a real testbed for multi-agent orchestration, AHB safety layers, and tool governance.
What can it actually do?
Out of the box, ask it to:
