Open Every Agent IDE — a self-hosted, front-end/back-end separated web IDE that unifies the best of today's agent surfaces: Codex-style agentic UX, Claude-style warmth, Muse-style voice embodiment, Grok-Bot-style scheduled automation, and opencode-style bring-your-own-provider flexibility. Four skins, one agent engine, any model.
Type a task in the browser; an agent plans it, runs real shell commands in a real workspace, edits real files, shows you diffs — with explicit approval gates and a full Review / Files / Terminal side panel. Every part is yours: your machine, your workspace, your API keys, your model.
browser (React + Vite) Node backend (Express)
┌───────────────────────────┐ REST + SSE ┌──────────────────────────────┐
│ sidebar · thread cards │ ◀───────────▶ │ agent loop │
│ composer (model/approval) │ │ plan → run_command → │
│ Review / Files / Terminal │ │ write_file → diff → reply │
│ settings · ⌘K palette │ ├──────────┬───────────────────┤
└───────────────────────────┘ │ providers│ local workspace │
│ OpenAI- │ git · fs · bash │
one `npm run dev` — no cloud, │ compat: │ (your machine) │
no account, no vendor lock-in │ Codex UI │ │
│ UX: │ DeepSeek, Qwen, │
│ plan & │ Kimi, GLM, Ollama │
│ cmd cards│ OpenRouter, … │
└──────────┴───────────────────┘
What "combines the strengths" means, honestly: the interaction surface follows the Codex desktop idea — threads with plan checklists,
$ commandcards, diff cards, approval pills, a Review panel — while the engine layer follows what opencode proved: any OpenAI-compatible provider, key stays on your machine, the agent is just a tool-calling loop. Open Every Agent IDE is an independent implementation built from scratch; it takes inspiration from publicly observed UX patterns of both ecosystems, and from neither copies code, assets, or branding.
Each product is a separate frontend + backend process with its own API surface, port and data directory — not a reskinned single app. Pick what to run at launch:
npm run dev:codex # IDE surface :5175 → :8787 ~/.open-every-agent/codex
npm run dev:claude # chat+artifacts:5176 → :8788 ~/.open-every-agent/claude
npm run dev:muse # studio+voice :5177 → :8789 ~/.open-every-agent/muse
npm run dev:grok # agent team :5178 → :8790 ~/.open-every-agent/grok| Product | Paradigm | API surface (guarded per product) |
|---|---|---|
| Codex | agentic IDE: sidebar threads, plan/command/diff cards, worktrees, checkpoints, steering, terminal, Review/Files/Canvas/Browser panels, ⌘K | full: fs/git/terminal/worktree/preview/processes/screenshot/open |
| Claude | chat-first: Recents sidebar, Projects, Artifacts grid + half-screen artifact panel (Preview/Code) | chats, projects, artifacts, approvals, search (fs/terminal blocked) |
| Muse | creative studio: icon rail, brief → creations, Gallery wall with live thumbnails, Boards, Voice embodiment, Cloud PC placeholder | boards, creations/artifacts, voice (fs/terminal blocked) |
| Grok Bot | agent team: roster with custom agents, Dispatch (multi-select members), per-agent transcript feed (swarm cards), Automations center, Cloud PC placeholder | tasks, agents CRUD, swarm dispatch, automations (fs blocked) |
All four speak to the same provider layer (13 OpenAI-compatible presets + custom) with per-product keys/config in their own state.json. Every surface is an original implementation of a publicly observed design language; no third-party code, artwork, fonts, icons, or logos are included.
| Product | Character (original SVG, state-machine animated) | Suggested provider |
|---|---|---|
| Codex | orbiting core + ring | any (DeepSeek/Qwen/…) |
| Claude | breathing ray-mark | Anthropic-compatible |
| Muse | egg that wobbles, cracks, hatches a creature | Meta Llama-compatible |
| Grok | blinking robot with antenna; bobs while working, smiles when done | xAI-compatible |
Mechanism notes from studying the real apps: Muse's animated character is a Rive state machine (idle → react → hatch) inside a 44 MB single-file web app; Grok Bot is a thin client for a cloud-rendered desktop agent. We reproduce the interaction idea — state-driven vector characters and a cloud-PC concept — with zero-dependency SVG/CSS and an honest Cloud PC placeholder panel (grok/muse skins): it simulates provisioning → boot → ready while the agent actually runs against your local workspace.
Every skin attaches any provider (13 presets incl. Anthropic, Meta, xAI, DeepSeek, Qwen, Kimi, GLM, OpenRouter, Groq, Ollama, custom). Settings → "SKIN → PROVIDER 映射" pins a default provider per skin — switching skins auto-switches providers, so each product surface can speak to its own vendor and to the shared external pool.
| Codex-style desktop agents | opencode | OpenEveryAgent IDE | |
|---|---|---|---|
| interface | native app, one vendor's account | terminal TUI | web UI in any browser |
| bring your own model | limited / vendor-bound | ✅ any provider | ✅ any OpenAI-compatible endpoint |
| plan / command / diff cards | ✅ | partial | ✅, streamed over SSE |
| approval gates | ✅ | ✅ | ✅ Auto / Ask / Read-only |
| real workspace (git, files, terminal) | ✅ | ✅ | ✅ with Review + commit UI |
| runs on | the vendor's binary | your terminal | npm run dev — laptop to server |
| multi-device access | no | hacky | it's a web app |
| license | proprietary | MIT-ish OSS | AGPL-3.0, source stays open |
What we do not claim: that this is a drop-in replacement for any hosted product, or that the built-in simulator matches a frontier model. Simulated mode exists so you can feel the full UX with zero keys; plug in DeepSeek, Qwen, Kimi, GLM, OpenRouter, Groq or local Ollama for real work.
git clone https://github.com/reynold-hu/OpenEveryAgentIDE.git
cd OpenEveryAgentIDE
npm run setup # installs backend + frontend deps
npm run dev # backend :8787 + frontend :5175
open http://localhost:5175Works with zero configuration (built-in simulator agent, real workspace, everything local). To connect a real model: Settings → Model Provider.
| Provider | Base URL | Example models |
|---|---|---|
| DeepSeek | https://api.deepseek.com/v1 |
deepseek-chat, deepseek-reasoner |
| Qwen (DashScope) | https://dashscope.aliyuncs.com/compatible-mode/v1 |
qwen3-coder-plus, qwen-max |
| Moonshot Kimi | https://api.moonshot.cn/v1 |
kimi-k2-0905-preview |
| Zhipu GLM | https://open.bigmodel.cn/api/paas/v4 |
glm-4.6 |
| OpenRouter | https://openrouter.ai/api/v1 |
any routed model id |
| Groq | https://api.groq.com/openai/v1 |
llama-3.3-70b-versatile |
| Ollama (local) | http://localhost:11434/v1 |
qwen3-coder, deepseek-r1 |
| OpenAI | https://api.openai.com/v1 |
gpt-5.1-codex |
| Custom | any /v1 endpoint |
free-form |
Paste the API key in Settings (kept only in ~/.open-codex-ide/state.json on
your machine — the API never returns it in plaintext, only hasKey: true).
Sync pulls the provider's live model list; or type a model directly.
Env fallback: OPEN_CODEX_API_KEY.
With a real provider, the agent runs a full tool-calling loop
(update_plan / run_command / write_file / read_file) streamed over SSE —
the same plan cards, command cards, diffs and approval gates as simulated mode,
with multi-turn context rebuilt from the stored transcript.
| Area | What you get |
|---|---|
| Shell | macOS-style titlebar, left nav rail (Home / Artifacts / History / Projects / Agents), collapsible sidebar, dark/light/system themes — deliberately logo-free & minimal in-app (the logo lives in this README + favicon) |
| Sidebar | project picker (+ create), search, chats grouped 置顶/Today/Yesterday/7-days, running-status pulse |
| Threads | SSE streaming, plan checklist cards, $ command cards with live output, merged "已编辑 N 个文件" card (+X −Y, 撤销, 查看变更, expandable file list), markdown replies |
| Thread menu (···) | 重命名 / 置顶 / 新建侧边聊天 / 继续 / Fork 到新对话 / 添加计划任务 / 分享 / 复制对话 / 新窗口打开 / 归档 |
| Summary panel | live task progress, changed files, web previews, command count, scheduled tasks, latest reply — every item jumps to its block in the thread |
| Web preview + headless browser | generated sites served from the workspace (/preview/:project/…), embedded preview pane, one-click Chrome headless screenshot, "打开方式" menu |
| Canvas | click any changed file to open it in an editable side canvas; save writes back to the workspace and updates the Review diff |
| Composer | model pill + approvals pill (Auto / Ask for approvals / Read-only), 📎 text-file attachments, Enter to send |
| Approvals | mutating commands & file changes pause for Allow/Deny cards; read-only mode blocks them |
| Side panel | Review (git status + colored diff + one-click commit), Files (tree + preview), Terminal (real shell in the workspace), 预览, 画布 |
| Messages | copy / read-aloud / 👍👎 feedback, scroll-to-bottom button, side-chat chips, ⌘S side chat |
| Scheduled tasks | recurring prompts that fire on the server even while the tab is closed |
| Settings | theme, provider + key, default model, approval mode, notification chime, reset |
| ⌘K palette | fuzzy jump to chats + New chat (⌘N), side chat, summary toggle, theme, settings, panel, new project |
| Persistence | chats/settings survive restarts; workspaces are real git repos under ~/.open-codex-ide/workspaces/ |
OpenEveryAgentIDE/
├── backend/ Node + Express REST API (port 8787)
│ └── src/
│ ├── index.js routes: settings / providers / models / projects / threads / approvals / fs / git / terminal / SSE
│ ├── agent.js agent loop: built-in simulator + real-LLM tool-calling engine (dispatch by provider)
│ ├── providers.js provider registry (DeepSeek/Qwen/Kimi/GLM/OpenRouter/Groq/Ollama/custom), key masking
│ ├── llm.js OpenAI-compatible streaming client + agent tool schemas
│ ├── workfs.js workspace fs, git status/diff/commit, shell exec
│ └── store.js JSON persistence (~/.open-codex-ide/)
└── frontend/ Vite + React + TypeScript (port 5175, /api proxied to 8787)
└── src/
├── App.tsx shell, state, SSE event folding, shortcuts
├── components/ Sidebar, Composer, Blocks (plan/command/diff cards),
│ SidePanel (Review/Files/Terminal), SettingsModal, CommandPalette
└── styles.css dark + light theme system
This is exactly the kind of project that wants native shells, and the
front/back split makes it easy: wrap frontend/dist in Tauri or
Electron and ship the backend/ as a sidecar process. Fork away, open an
issue to show off your port, and we'll link it here:
- Windows tray app (Tauri) — your fork could be this
- macOS menu bar + Dock (Codex-style) — or this
- iOS/Android via the web UI + hosted backend
Please keep the AGPL obligations and attribution intact (see below).
Licensed under AGPL-3.0-only (see LICENSE):
- ✅ Use, modify, and run it however you like — forking and building on it is encouraged.
- 🔗 Distribute or host a modified version (network service included) → you must publish the modified source under the same license and keep the copyright notice plus a link to this repo.
- 🚫 Authorship must stay visible: no stripping notices, no relabeling a derivative as your own closed product.
- ™ "OpenEveryAgent" is this project's name — forks should rename to avoid misrepresenting origin.
Codex is a trademark of OpenAI. opencode is a trademark of its respective owner. Claude is a trademark of Anthropic. Grok is a trademark of xAI. Muse is a trademark of Meta. Skin names here describe look-and-feel targets implemented entirely in original code — no assets, fonts, characters, or code were taken from any of them. This project is not affiliated with, endorsed by, or a continuation of any of these companies.
GET /api/settings | PUT /api/settings | POST /api/reset
GET /api/models GET /api/providers GET /api/providers/:id/models
GET|POST /api/projects DELETE /api/projects/:id
GET|POST /api/threads GET|PATCH|DELETE /api/threads/:id
POST /api/threads/:id/prompt GET /api/threads/:id/stream (SSE)
POST /api/approvals/:id
GET /api/projects/:id/tree|file|git POST /api/projects/:id/terminal|commit
SSE events: turn_started, plan, plan_update, command, command_update, command_result, approval_request, file_change, agent_message, agent_message_delta, agent_message_done, turn_complete, turn_error.