# Relay Coding Agent > Open-source desktop coding agent (Electron) with first-class support for non-mainstream/Chinese LLM providers — DeepSeek, Qwen, GLM, Kimi, MiniMax, and more. Chat + a full code workspace with file edits, scoped command execution, MCP plugins, and human-in-the-loop permissions. Early beta. - **Type:** MCP server - **Install:** `agentstack add mcp-leventenagy-relay-coding-agent` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [LeventeNagy](https://agentstack.voostack.com/s/leventenagy) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [LeventeNagy](https://github.com/LeventeNagy) - **Source:** https://github.com/LeventeNagy/relay-coding-agent ## Install ```sh agentstack add mcp-leventenagy-relay-coding-agent ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About Relay An open-source, dark-mode desktop coding agent — built for people who want to use non-mainstream LLM providers, not just the big three. English  ·  简体中文  ·  한국어 --- Relay is an Electron app that puts DeepSeek, Qwen, GLM, Kimi, MiniMax, and other open/Chinese models on equal footing with the usual suspects. Bring your own API key (or run locally via Ollama), pick a model, and work in either a **Chat** workspace or a full **Code** workspace with real file editing, command execution, and human-in-the-loop permissions. > ## 🌐 New: Browser use — the agent gets eyes *and* hands on the web > > Relay now has a **real browser built into the Code workspace** — one you drive *and* > the agent shares. **Log in once and stay logged in across restarts.** The agent can > **see** the page (screenshots + structured text) and **operate** it — navigate, click, > type, scroll — so it can work with sites, portals, and dashboards that have **no API**. > It can even start a dev server in the background and **open the app it just built** to > test it, closing the loop from *write → run → look → fix*. > > Just like the Claude desktop app, but open-source and wired into your coding agent. ## ⚠️ Project status — early beta (expect rough edges) **Relay is an early public beta (think alpha/0.1).** It already does a lot and is usable day to day, but it is *not* polished production software yet — you should expect bugs, half-finished corners, the occasional crash, and provider quirks (rate limits, model-specific errors, auth gotchas across the many supported providers). That's normal for software this young. **This is being shared openly precisely so we can fix those things together.** If you hit an error, please [open an issue](https://github.com/LeventeNagy/relay-coding-agent/issues) with what you did, which provider/model you used, and the error text or a screenshot — that's the single most helpful thing you can do. Bug reports, fixes, and ideas are all genuinely welcome; with more eyes on it we can squash these issues much faster than I can alone. See [CONTRIBUTING.md](CONTRIBUTING.md). Known in-progress areas: signed installers / auto-update, broader automated test coverage, and additional renderer hardening. Your API keys are **not** affected by any of this — they're encrypted on your own machine and never leave it (see the security model below). 📣 **Follow [@_levyathan_](https://x.com/_levyathan_) on Twitter/X** for updates, progress, and release notes. ## Highlights - **Many providers, one router.** DeepSeek, Alibaba/Qwen (DashScope + coding/token plans), Z.AI / GLM, Moonshot / Kimi (global + China), MiniMax, SiliconFlow, Tencent, Xiaomi, OpenRouter, and Ollama Cloud — via the Mastra model router. Add a key in Settings and its models appear in the picker. - **Chat workspace.** Streaming replies, reasoning ("thinking") controls where the model supports them, image/document attachments, and web **search** + deep **research** (Tavily / Brave / DuckDuckGo, key optional). - **Code workspace.** Projects live in real folders. The agent reads, writes, and edits files and runs commands **scoped to the project**, with permission modes (Ask / Approve / Full). **Plan mode** asks clarifying questions and proposes a plan before touching anything. **Sources** (your design docs + framework docs) are kept in the agent's context every turn so it never loses the plot, and new projects default to a **Next.js + shadcn/ui** stack the agent scaffolds and keeps current. - **🌐 Browser use.** A persistent, embedded browser in the Code workspace that you and the agent **share**. Logins survive restarts; the agent perceives pages via screenshots + structured text and drives them (navigate/click/type/scroll) — great for sites with no API. It can launch a dev server in the background and open the app it just built to test it. - **Plugins (MCP).** Connect Model Context Protocol servers — Notion and Linear via one-click OAuth; GitHub, Filesystem, Git, Supabase (OAuth), Convex, Postgres, and more. Plugins are chosen per conversation and filtered by workspace. - **Skills.** Reusable instruction packs the agent can pull in on demand. - **Parallel sessions.** Start a run in one session and switch away — it keeps streaming and saving in the background, with a "working" indicator. - **Context management.** Per-model context windows, automatic compaction of older turns, and a live usage meter. ## Security model (short version) - API keys, plugin credentials, OAuth tokens, and search keys are encrypted at rest with Electron `safeStorage` in the app's userData directory. - Decrypted secrets never leave the main process — the renderer only ever sees key *names* and connection status. - File and command tools are confined to the active project folder; writes and shell commands are gated by the permission mode (path traversal is blocked unless you grant Full access). - The renderer runs with `contextIsolation` and no Node integration; links in chat open in your real browser. The opt-in **shared browser** is a separate, sandboxed session with its own storage partition, kept isolated from the app's own renderer. See [SECURITY.md](SECURITY.md) for the full threat model and how to report issues. ## Getting started Prerequisites: **Node.js ≥ 20** and **Corepack** (bundled with Node). Local (stdio) MCP plugins like Filesystem/Git also need `npx` on your PATH. ```bash corepack pnpm install # install dependencies (pnpm 11) corepack pnpm dev # run the app in development corepack pnpm build # typecheck + production build into out/ ``` Then open **Settings**, add an API key for any supported provider, and pick a model. ## Tech stack Electron · electron-vite (Vite) · React 19 · TypeScript · [Mastra](https://mastra.ai) (`@mastra/core` model router + `@mastra/mcp`). ## Contributing Issues and PRs are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md). Run `corepack pnpm typecheck && corepack pnpm lint && corepack pnpm test` before opening a PR. ## License [MIT](LICENSE) © Levente Nagy ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [LeventeNagy](https://github.com/LeventeNagy) - **Source:** [LeventeNagy/relay-coding-agent](https://github.com/LeventeNagy/relay-coding-agent) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-leventenagy-relay-coding-agent - Seller: https://agentstack.voostack.com/s/leventenagy - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.