So, I have tried lots of the popular local agent frameworks. Hermes Agent, OpenClaw, Agent Zero and so on. It was great fun and extremely educational, but I felt that none of them are ready for most end-users. They can take forever to setup, or simply break when you finally spent weeks on training them. So I decided to build my own!

Fixes several issues with current frameworks

Lots of things can go bad when working with agents, and I have tried to address them all in my own agent which hopefully will be publicly available soon. Here are some of the improvements over competing frameworks…

– Built with PHP for Linux systems
– Simple installer with just a few questions
– Easy LLM setup that just works (both local for GPU, cloud provider and tiny CPU backup)
– View and handle local models in Ollama directly from the graphical interface
– WebUI with password protection
– Clever long-term memory with markdown Wiki and QMD semantic search
– Core files are separated from the actual workspace for easier backup or restore
– Agent by default doesn’t have permission to edit (or crash) the agent core
– Lots of guardrails that prevent looped thinking or stalled work
– Pre-built subagents optimized for common tasks
– Consultant subagent that spends tokens on external LLM provider only when needed
– Agent overview to monitor live progress and delegations
– Quick setup of LAN or internet access to the agent
– Progressive Web App; works on Android without involving Google Play
– Kanban project management board that even idiots understand to use
– File explorer where the entire Wiki and workfiles can be viewed
– Skills for better long running and multi-step challenges
– MCP for adding endless capabilities by external services
– Truly automatic context window handling
– A single continuous chat per project, no more session switching
– Toggle setting for giving the agent full system access or sudo

Some improvements, debugging and testing is still underway. But keep an eye out for Dagento!