// the take
Minke is an Electron desktop shell around DeepSeek Harness (DSH) that bundles chat, file management, a terminal, and a shared/controllable browser into one agent workspace, with remote access from a phone via Tailscale, Cloudflare Access, or chat bots (WeChat/Telegram/Discord). It's aimed at developers who already use DSH-style coding agents and want a GUI plus remote control, not at people looking for a standalone agent framework.
The shared browser feature is a real differentiator — you can watch the agent drive a visible tab, take over mid-task, and feed annotated page context back into the conversation, which is genuinely useful for debugging agent browsing behavior. The remote access design is unusually careful for a side project: Tailscale Serve keeps Harness on loopback with no LAN binding and no public Funnel, and the docs are explicit about the security tradeoffs of each transport (HTTPS Serve vs. plain-HTTP direct IP vs. Cloudflare Access). The auto-update flow verifies the full chain (fixed asset per platform, exact size, SHA-256, GitHub release URL, OS provenance marker) before prompting to install, which is more rigorous than most Electron apps bother with.
It's a thin, tightly-coupled wrapper around a pinned pre-release of DeepSeek Harness (v0.2.1-rc.2) pulled in as a git submodule with custom patches (cordis.patch.yml) — you inherit whatever bugs or security issues exist in an RC build of someone else's agent runtime, and upgrading means re-applying patches by hand per their own upgrade notes. The remote surface is broad: once a device is on the tailnet or passes Cloudflare Access, the web client can start agent tasks and use any local tool already authorized in Minke, so a single access-control mistake on Tailscale/Cloudflare hands out agent-level control of the host machine, not just read access. macOS pre-release builds aren't notarized, and the docs walk through re-signing the app with an ad-hoc signature (`codesign --force --deep --sign -`) as a fix for Keychain access failures — that quietly undermines the code-signing identity that macOS Keychain access and credential storage are supposed to depend on. There's no cross-platform build from a single host, so producing all five package types requires three separate machines/CI runners, and the project is explicitly a single-maintainer community effort still in active flux (packaged releases may diverge from source, per their own disclaimer).
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →