Agent CLIs are evolving fast—tools keep piling up and it is easy to lose track.
Some say the CLI is the future?
This is MindFast’s core design principle.
Usually, to support widgets or artifacts, you give the agent bespoke tools—how to create a card, render a chart, or update state. MindFast does not take that path.
A capable agent stack, in theory, only needs four tools: Read, Write, Edit, and Bash (sometimes even just Bash). Extra tool definitions add complexity and constrain what the agent can do. Instead of a custom protocol for every UI element, the agent does what it does best: work with files.
MindFast therefore uses directory conventions:
- Files under
widgets/render as small components in the chat stream - Files under
artifacts/render as resources in the sidebar - Files under
plans/track the plan for the current conversation
The agent does not need new tool shapes—only which directory to write to. Rules for widgets and artifacts can live in the system prompt or in Guideline files loaded on demand.
The upside is clear: when a plan step completes, Edit the right line in the right file; to show a chart, drop an HTML file into widgets/. The agent keeps doing what it is good at: reading and writing files.
In the current web build, MindFast uses IndexedDB to emulate full file create/read/write, and the Bash tool emulates common shell commands. That layer supports the four tools above and sets the stage for a future Tauri build with a real shell. Moving from the browser sandbox to the local machine should be smooth without rewriting the upper layers.
Interactive components embedded directly in the chat. Charts, forms, code previews, plan boards—generated by the agent—show up in context, not only in a separate panel.
Persistent resources in the sidebar. Better for larger outputs: full source files, long documents, multi-step write-ups. Unlike widgets, artifacts sit outside the message stream so you can open and compare them anytime.
Rules and conventions loaded on demand. Instead of stuffing everything into the system prompt, split scenarios into separate Guideline files the agent pulls when needed. That cuts context noise and keeps policies easier to maintain.
For complex tasks, MindFast can maintain a plan and track progress step by step.
Sub-agents spawned via Bash commands. Each SubAgent is still a full agent loop: instructions, tools, results. Today the main agent starts them through Bash; they run in isolated context and return output to the parent.
That lets heavy work split into parallel subtasks without crowding one conversation. Deeper orchestration, dependencies, and cross-agent messaging are future work; the current focus is a solid single level of delegation.
MindFast runs as a web app today and is planned to ship as a Tauri desktop client, enabling:
- Real local filesystem access
- Native shell execution
- MCP
- Skills
- Agent-to-agent communication
The architecture is already shaped so the jump from virtual to real stays incremental.
Issues and suggestions are welcome.
MIT © Solaren Zhou