DeskPilot: open source native Python desktop agent client for local LLMs with MCP and sandboxing

poofph · reddit · 2026-09-24

A developer released DeskPilot, an open source native Python/Tkinter desktop agentic client for local LLMs (built with help from Qwen 3.8 27B).

Highlights:

Universal backends: works with Ollama, LM Studio, vLLM, llama.cpp, or any OpenAI-compatible endpoint
MCP integration: dependency-free stdio JSON-RPC 2.0 client that dynamically converts MCP tools into function calling schemas
Permission bar: per-tool Always Allow / Ask Permission / Off controls, with unvetted tools defaulting to ask
Built-in tools: SearXNG/Exa search, Firecrawl scraping, file/PDF I/O, clipboard, screen capture, sandboxed Node.js execution
Reasoning support: streaming <think>/reasoningcontent display, auto-collapsing to "Thought for X.Xs"
Fully local voice stack: faster-whisper STT + kokoro-onnx TTS with lookahead pipeline
Context management: real decode tok/s, TTFT, live context gauge, and automatic structured handoff summaries at a user-defined threshold (e.g. 75%)
Security: SSRF/DNS rebinding protection, filesystem sandboxing, and daemon threading for a fluid UI

Runs on Python 3.10+ or compiles to a single .exe via PyInstaller.

More from coding & agent