Comment by manukminasyan
Hi HN, I'm Manuk. I've been working on Relaticle for about two years, and I built most of it myself.I started building Relaticle as a traditional CRM tool but when I added AI agents I had to decide how much access to give them. On one hand, read-only access had limitations for some cases, on the other hand, allowing the agents to make changes without asking did not feel right.
So I decided to separate 2 use cases. External MCP clients use OAuth and they can make changes directly, just like API clients.
The assistant in CRM works in a different way. First it shows what it wants to change, then the user either approves or rejects it. For batch updates, you can review the entire record, approve and keep what is needed.
The MCP server now has 37 tools for companies, people, opportunities, tasks, notes, search and relationships. Each workspace can have its own custom fields and these fields are automatically available in both MCP and chat.
Later I had to rebuild the chat as well, because of some issues, for example a message could disappear when a retry was stopped, or sometimes a draft could end up in a wrong conversation. When rebuilding I fixed theses issues and added saved drafts, switches between conversations,loads history as you scroll as well as records links in answers.
You can run the CRM, MCP server and the assistant on your own server. Ollama keeps the model interface on your infrastructure. The repo includes Docker Compose and setup guides, so you can run everything without creating a hosted account.
The stack is PHP 8.5, Laravel 13, Filament 5, Livewire 4, PostgreSQL, Redis, Horizon, and Reverb.
One thing I am still thinking about: should an assistant be allowed to make auto-approvals based on certain rules or is it better to ask for user approval every time.