5,876 Commits Across Three AI Products

5,876 commits. Three repositories. Four years and four months.

That’s the current state of what we now call LikeClaw – a platform that started as a forked workflow automation tool in 2021 and, through a series of pivots, rewrites, and stubborn engineering decisions, became an AI agent platform with sandboxed code execution.

This is the story of building three AI products, what each one taught us, and why we’re still building.

October 2021: A fork and a bet

On October 1, 2021, we forked Automatisch, an open-source Zapier alternative. The initial commit was chore: Add monorepo setup. Not dramatic. Not visionary. Just a monorepo config and a bet that businesses would want self-hosted workflow automation.

The first year was integration work. Slack, Twitter, Google Sheets, Twilio, Discord, Telegram, Notion, Trello. One integration at a time, each with its own OAuth flow, its own API quirks, its own edge cases. By the end of 2022, we had 1,069 commits and a functional workflow automation platform that could chain triggers and actions across dozens of services.

We called it Autopilot. The selling point was simple: your data stays on your servers. For companies in healthcare, finance, or any industry bound by GDPR, that mattered more than Zapier’s 8,000-integration catalog.

What Autopilot taught us: Integration breadth is a moat, but it’s a slow moat. Every plugin is a maintenance commitment. We now maintain 168 integrations, and every API deprecation notice from a third-party service creates work.

February 2023: The pivot nobody planned

On February 3, 2023, we created aiwayz-dashboard-api. The first commits were for a news aggregation service – fetch RSS feeds, process them with OpenAI, archive to Google Cloud Storage. We called it “AI News.”

Within two months, the news aggregator was dead. By April, we had a chat feature running on GPT-3.5. By May, we had research agents that could plan and execute multi-step tasks with Google Search. The project got renamed to “dashboard-api.”

The pivot happened because we kept asking ourselves: what if the AI isn’t just processing data in the background? What if it’s the interface?

From there, the feature list grew fast. Document composition with 100+ templates. Knowledge bases with vector search. Custom chatbots. Image generation with Stable Diffusion and DALL-E. Claude integration. Pro Search. Organization management with role-based access. By the end of 2024, we had 2,892 commits, 31+ MongoDB collections, and a multi-model AI dashboard supporting GPT-4, Claude, Gemini, DeepSeek, and 20+ models through OpenRouter.

What Dashboard v1 taught us: Multi-model access isn’t a feature. It’s a requirement. Our users didn’t want to pick one model and stick with it. They wanted Claude for writing, GPT for code, Gemini for analysis, and the flexibility to switch without juggling four subscriptions.

The part nobody talks about: infrastructure

While features got the attention, the real work happened in the infrastructure layer. We built:

• A centralized auth service with JWT tokens shared across all three products
• A billing service on Stripe that evolved from direct integration to an external API to a credits-based pay-as-you-go system
• RabbitMQ for inter-service communication (five handler types: wizard, approval, users, organizations, management)
• Three deployment environments per product: staging, production-global, production-Russia
• 18+ GitHub Actions workflows for CI/CD
• Pulumi infrastructure-as-code for Kubernetes on GKE
• Langfuse for tracking AI token costs across every model and every request

None of this is exciting. All of it was necessary. Every time we added a product, the shared infrastructure had to scale with it. The billing system alone was rewritten three times.

November 21, 2025: The rewrite

We started Dashboard v2 on November 21, 2025. A complete rewrite. NestJS 11, React 19, Tailwind 4. No migration path. No backward compatibility. Clean slate.

The reason was simple: we’d outgrown the original architecture. The Koa-based API with its 120+ endpoints and 31+ MongoDB collections worked, but it wasn’t designed for what we needed next – agents that could execute code, manage files, and operate autonomously in sandboxed environments.

By day nine, we had chat, a virtual file system, workspaces, internationalization in four languages, and file attachments. The pace was reckless by any normal measure. 163 commits in the first ten days.

In 88 days, we shipped:

• 40+ specialized AI agents, each with distinct capabilities and personalities
• E2B sandboxed code execution with three sandbox templates, SSH authentication, and bi-directional file sync
• A virtual file system backed by MongoDB with automatic version snapshots before every mutation
• A skills marketplace with SKILL.md format, ClawHub import, and per-user approval
• Background task delegation for long-running operations with smart chaining and loop detection
• A scheduling system with cron-based execution and lock-based duplicate prevention
• An evaluation framework with LLM-as-judge scoring and GitHub Pages reporting

The commit log tells the story. feat: add anti-freeze watchdog and OpenRouter provider routing. feat: smart background task chaining with loop detection. feat: add VFS file versioning with history, restore, and task rollback. Each commit is a problem we didn’t know we’d have until we hit it.

What Dashboard v2 taught us: The agent paradigm changes everything. When your AI can execute code in a sandbox, manage files, browse the web, and delegate tasks to other agents, the product isn’t a chat interface anymore. It’s an operating system.

February 2026: Three products become one

On February 16, 2026 – four years and four months after the first Autopilot commit – we shipped the LikeClaw variant of Dashboard v2. Local authentication, Stripe credits, OpenRouter LLM proxy, Cloudflare Pages deployment. A self-contained product that bundles everything we’d learned.

The three products still exist as separate codebases. Autopilot handles event-driven workflows. Dashboard v1 serves existing users with its mature feature set. Dashboard v2 powers the agent platform. But they share infrastructure, and they’re converging.

Together, they represent:

• 5,876 commits from a team that never stopped shipping
• 168 integration plugins across communication, social media, productivity, CRM, AI, and data tools
• 30+ AI models from OpenAI, Anthropic, Google, and OpenRouter
• 40+ specialized agents with sandboxed code execution
• 4 languages (English, Russian, Chinese Simplified, Chinese Traditional)
• 85 PostgreSQL migrations in Autopilot and 31+ MongoDB collections in the dashboard

What we’d tell ourselves four years ago

Don’t optimize for the architecture you have. Optimize for the speed at which you can change it. We’ve swapped databases, rewritten billing systems, migrated from streaming to polling, replaced Pinecone with SmartFAQ, moved from .skill files to SKILL.md directories, and killed features we spent months building. The code that survives isn’t the most elegant. It’s the most adaptable.

Every product you build teaches you what the next product should be. Autopilot taught us that automation without AI is just plumbing. Dashboard v1 taught us that multi-model AI without code execution is just a chat window. Dashboard v2 taught us that agents without sandboxing are a liability.

Ship the billing system first. We wrote about this in another post, but it bears repeating. Credits-based pay-as-you-go was the right model all along. We just needed three iterations to get there.

The next four years will look nothing like the last four. The agentic AI market is projected to grow from $7.8 billion to over $50 billion by 2030. Multi-agent systems had a 1,445% surge in enterprise inquiries in the past year. We’re not building for a niche anymore.

5,876 commits in. Still shipping.