Your AI workspace lives in your browser.
Live demo: tryaidiy.com
Local-first, bring-your-own-key chat for Node or Docker. Your workspace state stays in the browser by default, while the server relays requests to the model and tools you choose. Provider API keys are not required as persistent server configuration.
Built with React Router, assistant-ui, the Vercel AI SDK, Tailwind CSS, browser-side Pyodide, and CheerpX (in-browser Linux).
Demo— open the app and try the workspace** Deploy**— one command (Node or Docker Compose)** GitHub**—Cubinghackerz/ai.diy
npm install
npm run build && npm start
Open http://localhost:3000
→ Open workspace → add your API key → chat.
docker compose up --build
docker build -t ai-diy .
docker run -p 3000:3000 ai-diy
npm run dev
is not recommended (composer regression in Vite). Use the production build above. Details: DEPLOYMENT.md · QA checklist: QA.md · launch brief: PRODUCT_HUNT.md.
Features marked available are wired. Planned items are not claimed as working.
Available: landing, 20+ provider integrations, model discovery, local persistence, files, browser Python (Canvas capture + IndexedDB persistence for generated images/binaries), in-browser Linux environment (CheerpX/WebVM: bash, python3, gcc, node, apt; no outbound network by default), search + connectors, remote MCP, artifacts, memory, on-device knowledge RAG, usage ledger with soft spend/token/RPM caps, server rate-limit hooks, voice dictation (Web Speech), multi-model Preview, import/export, client-side S3/WebDAV/Google Drive backup, portable skills catalog + install, slash commands (/Research
,/Compaction
,/Subagent
, …), Agent Mode, subagents (approve → wait → synthesize), encrypted browser settings where supported, Vercel Connect (Beta: token-backed MCP servers +connect_request
), first-run and Settings tool-access allowlist (only enabled capabilities are registered for the model)Browser-local npm projects: the AI can scaffold a per-chat Node project in the active browser tab using WebContainers, write files, install public registry packages with lifecycle scripts disabled, run allowlisted build/test scripts, inspect files, and export a tarball. Projects run in the browser, not on the ai.diy server or CheerpX Linux VM.Coming soon: direct GitHub/Supabase/PostgreSQL adapters, custom-provider capability probing
| Component | Location |
|---|---|
| Chat UI, settings, history, Canvas artifacts, memory, knowledge base, usage events, Preview | Browser (localStorage + IndexedDB) |
| Dictation, Python, and Linux VM | Browser (Web Speech + Pyodide + CheerpX) |
| LLM relay, model discovery, search, MCP, optional RPM rate limit | Node server |
| Provider API keys | Browser only; relayed per request |
When Web Crypto and IndexedDB are available, settings are encrypted at rest with AES-GCM. The encrypted payload is kept in localStorage and its envelope key is kept separately in IndexedDB. Browser or platform fallbacks may use plaintext storage, and this protection does not defend against a compromised browser profile or malicious same-origin code. Provider keys still pass through the relay in transit; treat hosted instances as able to observe that traffic.
Browser-owned workspace
settings, provider keys, chats, memory, knowledge, Canvas, Preview
|
| per-request relay
v
Node server
model forwarding, model discovery, optional search/MCP/connectors
|
+--> selected cloud or local model provider
+--> explicitly enabled search, fetch, MCP, or connector service
The server does not need persistent provider API keys. It can still see request traffic while relaying it, and enabled tools can send selected data to their own services. See ARCHITECTURE.md and SECURITY.md.
Portable skill packages live in skills/. Install from
Settings → Skills(search → Install → available via
/
in the composer).Built-in slash commands force matching tools for the next send — including ** /Linux Environment** (in-browser Debian VM),
(spawns approved subagents; the main chat waits for results before continuing),
/Subagent
/Research
, /Compaction
, and design/file-creation skills. Skill tool calls appear in chat as Used skill: ….
When a request asks for an npm-backed app or Node library, ai.diy activates NPM Project. The workflow stays inside a browser-native WebContainer: initialize a project, write its files, install exact public registry packages, run build
, dev
, start
, preview
, test
, lint
, typecheck
, check
, or format
, then read or export the result. It does not depend on the Debian Node 10 CheerpX VM.
Flagship starters include Deep Research, Code Review, GitHub Repository Analysis, PDF Analysis, Incident Investigator, and General Task Solver (understand → select skills → execute → verify → synthesize).
Authoring guide for agents: .cursor/skills/ai-diy-skill-authoring/SKILL.md.
Enable Agent Mode under Settings → Experimental. The model plans, selects installed skills/tools, verifies, and synthesizes. Pair with General Task Solver.
Subagents (same Experimental section, or force with /Subagent
): the model calls spawn_subagent
/ spawn_subagents
; you approve each run in a popup; the main chat s until those sessions finish (or are declined/cancelled), then continues from structured results. Nested subagents cannot spawn further subagents or ask the user questions.
Settings → Knowledge Base uploads documents for on-device RAG (browser WASM embeddings + HNSW). The index stays local and is not uploaded to a vendor vector store, but retrieved context may be sent to the selected cloud model. The model can call knowledge_search
/ knowledge_list
.
Settings → Usage & cost shows provider-reported tokens/cost and optional soft caps (daily spend, tokens, RPM) keyed by API-key fingerprint in IndexedDB. The server also supports sliding-window RPM limits via RATE_LIMIT_RPM
/ RATE_LIMIT_DISABLED
(see .env.example
). Soft caps are client-enforced; configure server limits before public exposure.
OpenAI, ChatGPT subscription, Anthropic, Gemini, Groq, Cerebras, Fireworks, Perplexity, Cohere, OpenRouter, xAI, DeepSeek, Bedrock, Azure, Vertex, Vercel Gateway, Together, Mistral, Hugging Face, Ollama, LM Studio, and custom OpenAI-compatible endpoints.
Web search (DuckDuckGo + connectors), URL fetch, calculator, browser Python, in-browser Linux (linux_run_command
/ linux_read_file
), files/artifacts, research and design skills, local time, memory, knowledge search, ask user, remote MCP (Firecrawl + Parallel bundled keyless), subagents.
Python saves generated files in the working directory; the browser captures up to four files (≤2 MiB each) into Canvas and persists them with the chat when under the client size cap.
The Linux environment is a client-side Debian VM (CheerpX/WebVM) with bash, python3, gcc, node, and apt. Enable it under Settings → Linux environment. Networking is offline by default; raw Linux apt/pip/npm/curl/git downloads require the opt-in Tailscale bridge and an internet exit node. Use NPM Project for browser-native npm work without that VM network. The VM stays warm in the tab after the first boot. Commands time out at 90s. /workspace
is cross-origin isolated (required for SharedArrayBuffer). While a reply or VM command is running, send and new-chat are locked; Stop aborts the wait and unlocks the composer. A corrupt disk overlay is discarded and the VM retries once instead of hanging.
Global custom instructions live under Settings → Instructions. They append to ai.diy's defaults rather than replacing core tool, safety, and active-skill instructions.
Token mode (Settings → Token mode, or the token/TTFT chip above any assistant message) controls system-prompt size, tool suite, step budget, and prompt caching: Efficient, Balanced (default), Prompt caching, and Full suite.
Outside Full suite, optional capabilities are registered compactly and their full instructions are served on demand: the model calls load_tool_guide
and reads the exact guide for Python, files, URL Doctor, knowledge, memory, skills, and subagents only when a turn actually needs them. Enabled skills appear in the on-demand catalog by name instead of being injected into every prompt. Web search and URL fetch stay always available; the bundled Firecrawl/Parallel MCP schemas are only discovered on turns that ask for live web research.
ChatGPT, Claude, ShareGPT, Markdown, ai.diy JSON. Client-side backup to S3-compatible storage, WebDAV, or Google Drive (service-account JSON). Credentials stay in the browser and talk only to your storage endpoint.
Note: an older “Cloud storage coming soon” blurb referred to Google Drive
OAuth. Client-side S3/WebDAV/Drive service-account backup is available now.
- No LLM keys in server env
- SSRF guards on
fetch_url
and private provider/MCP URLs (unlessALLOW_PRIVATE_PROVIDER_URLS=true
on trusted self-host) - Stdio MCP rejected
- Optional server
RATE_LIMIT_RPM
(enable before public exposure; see.env.example
) - Client soft usage caps in Settings → Usage & cost
- Do not log request bodies or credentials
- In-browser Linux (CheerpX) runs entirely in the tab: Safari is supported; there is no outbound network. A corrupt disk overlay is wiped and retried automatically. The CheerpX runtime is loaded from Leaning Technologies' CDN under the
CheerpX Community License(free for FOSS such as this MIT project). Commercial use or self-hosting the runtime requires a license from Leaning Technologies — do not vendor
cx.esm.js
.
See DEPLOYMENT.md, SECURITY.md, and .env.example
.
npx vercel
Use this for preview testing. For a public launch, configure a branded domain
and set VITE_SITE_URL
at build time before using a production deployment.
ai.diy stands on a lot of excellent open work:
assistant-uiandVercel AI SDKReact RouterandTailwind CSSCheerpX / WebVMby Leaning TechnologiesPyodideLogin with ChatGPTcmdk,Base UI,Radix UI,Phosphor Icons,Lucidemodels.devfor model catalog dataFirecrawlandParallelfor the bundled keyless search MCPsxterm.jsfor the in-browser terminal
MIT. See LICENSE.