The Pitch: Escape the OpenAI-Anthropic-Google Trap

Relay is a new open-source desktop coding agent that challenges the assumption you must use one of the big three providers to get serious AI-assisted development work done. Built by developer Levente Nagy and released as early beta, Relay is an Electron app designed from the ground up for people who want to run DeepSeek, Alibaba's Qwen, Z.AI/GLM, Moonshot/Kimi, MiniMax, or a dozen other open-weight and Chinese-origin models without jumping through hoops. The pitch is simple: bring your own API key (or go fully local via Ollama), pick your model, and get to work in either a Chat workspace for exploration or a full Code workspace with real file editing, command execution, and human-in-the-loop permission controls.

Provider Support Runs Deep

The provider list is where Relay differentiates itself most sharply. Beyond the obvious alternatives, Nagy has integrated DashScope (Qwen's coding and token plans), SiliconFlow, Tencent Cloud, Xiaomi's AI offerings, OpenRouter, and Ollama Cloud through Mastra's model router. This isn't a half-baked dropdown with three options — it's a genuine multi-provider architecture that treats all these models as first-class citizens. The practical implication: if you've been locked into Claude or GPT-4o because the ecosystem made it easy, Relay makes that same flexibility available for whatever model fits your use case, budget, or geopolitical preferences.

Chat vs. Code Workspaces Serve Different Workflows

Relay splits functionality into two modes. The Chat workspace handles conversational interactions with streaming replies, reasoning/thinking controls (where supported by the model), image and document attachments, and web search via Tavily, Brave Search, or DuckDuckGo — keys optional for most. The Code workspace is where things get interesting for builders: projects live in real filesystem folders, the agent can read, write, and edit files directly, execute shell commands scoped to the project directory, and operate under three permission tiers (Ask/Approve/Full). Plan mode deserves special mention — it lets the agent ask clarifying questions and propose a modification plan before touching anything, which is exactly the kind of guardrail power users want when handing a coding assistant write access to their repo. New projects scaffold with Next.js and shadcn/ui by default, keeping that stack current as work progresses.

Security Model Is Worth Examining

Nagy's security architecture shows careful thinking. API keys, plugin credentials, OAuth tokens, and search keys are encrypted at rest using Electron's safeStorage API in the app's userData directory — decrypted secrets never leave the main process, so the renderer only sees key names and connection status. File and command tools are sandboxed to the active project folder, path traversal is blocked unless you grant Full access, and the renderer runs with contextIsolation enabled and Node integration disabled. Links open in your real browser rather than an embedded WebView. For a tool handling API keys for a dozen different providers across Chinese, American, and European companies, this level of hardening matters — especially if you're running models from multiple sources simultaneously.

Status: Alpha-Grade Software With Real Potential

The README doesn't sugarcoat it: Relay is early beta (Nagy describes it as "alpha/0.1" equivalent). Users should expect bugs, crashes, provider-specific quirks like rate limiting and auth gotchas, and rough UI corners. In-progress work includes signed installers with auto-update, broader automated test coverage, and renderer hardening improvements. Nagy explicitly asks for bug reports with reproduction steps, provider/model info, and error text or screenshots — the classic open-source call to arms that works when the community shows up. The tech stack (Electron + electron-vite + React 19 + TypeScript + Mastra) is modern and maintainable, which suggests long-term viability if contributor interest materializes.

Key Takeaways

  • Relay supports 12+ non-mainstream LLM providers including DeepSeek, Qwen, GLM, Kimi, MiniMax, SiliconFlow, Tencent, Xiaomi, OpenRouter, and Ollama Cloud
  • Code workspace offers real file editing, scoped shell commands, and a Plan mode that proposes changes before execution
  • API keys stay local via Electron safeStorage encryption; renderer isolation prevents key exposure even if the UI is compromised
  • Early beta means rough edges — contributors and bug reporters will shape the project's trajectory

The Bottom Line

Relay fills a gap that's been widening as the AI provider landscape fragments. If you've been watching DeepSeek's coding performance or Qwen's multilingual capabilities with interest but couldn't find tooling that treated them seriously, this is your entry point. Just don't mistake early beta for production-ready software — treat it like any other alpha project, file those bug reports, and you might help shape something genuinely useful for the hacker crowd that's tired of corporate AI monoculture.