Molf AssistantSelf-hosted AI agent

A client-server-worker architecture for running LLM agents with local tool execution.

Client-Server-Worker Architecture

A central tRPC WebSocket server orchestrates LLM interactions. Workers execute tools locally. Multiple clients (TUI, Telegram, custom) connect over TLS-secured WebSocket.

16 LLM Providers

Gemini, Anthropic, OpenAI, and 13 more via Vercel AI SDK with automatic API key detection. Switch models per-workspace or per-prompt. Dynamic model catalog from models.dev.

TLS with TOFU Trust

TLS enabled by default with auto-generated certificates. Trust-on-first-use model for easy setup -- approve the fingerprint once, and it's pinned for future connections.

Built-in Tools & Skills

Six built-in tools (shell, file read/write/edit, grep, glob), lazy-loaded skill documents, and MCP integration for connecting external tool servers.

Subagents

Spawn isolated child agents for parallel or specialized subtasks. Two built-in agents (explore, general) plus custom agent definitions via Markdown files.

Tool Approval & Plugins

Configurable per-tool approval rules with glob pattern matching. Extensible plugin system with 15 server hooks and 4 worker hooks.

Session Persistence

Sessions stored as JSON files with automatic context summarization and pruning. Survive restarts and worker reconnections.