Deep Research Web UI

[English | 中文]

Deep Research Web UI turns a research question into a cited report: it plans searches, browses the web, verifies findings against real source pages, and lets you inspect and follow up on every citation — all from your browser, or self-hosted with your own API keys.

Features:

• 🚀 Safe & Secure: In Client Mode, config and API requests stay in your browser locally
• 🕙 Realtime feedback: Stream AI responses and reflect on the UI in real-time
• 🌳 Search visualization: Shows the research process using a tree structure. Supports searching in different languages
• 📄 Export as PDF: Export the final research report as Markdown / PDF
• 🤖 Supports more models: Uses plain prompts instead of newer, less widely supported features like Structured Outputs. This ensures to work with more providers that haven't caught up with the latest OpenAI capabilities.
• 🐳 Docker support: Deploy in your environment in one-line command
• 🔧 Server Mode: Deploy with environment variables, no need for users to configure API keys
• 🔎 Evidence and follow-ups: Inspect source excerpts from report citations, then research a finding further and update related passages while keeping the original report in history
• 🎯 Focused search: Separate queries from time and source filters, check result relevance, and retry once with a revised query when evidence is insufficient
• 📖 On-demand full-page reading: Fetches full source pages to extract verbatim evidence when search snippets aren't enough, sharing a token/request budget across the whole research run

Currently available providers:

• AI: OpenAI compatible, ApiSmart, SiliconFlow, DeepSeek, OpenRouter, Requesty, Ollama, LiteLLM and more
• Web Search: Tavily (1000 free credits / month), Firecrawl (cloud / self-hosted), fastCRW (cloud / self-hosted), Google PSE, You.com, Serply, Parallel (keyless, Server Mode only)

Please give a 🌟 Star if you like this project!

---

Sponsors

MangoProxy provides Residential, ISP, Mobile, and Datacenter proxies in 200+ locations for web scraping, automation, SEO, and multi-account management. Promo code: GITHUBISP - 8% off Static ISP proxies.

---

ApiSmart.ai provides unified access to leading AI models through a single API. Use one API key to connect with LLM, image, and video models through an OpenAI-compatible interface, without managing multiple providers separately. Switch models easily, simplify billing, and improve reliability with intelligent routing and automatic failover. Build and scale AI applications faster with one streamlined API platform.

---

Helodata provides global proxy infrastructure covering 195+ countries and regions, with 80M+ ethically sourced residential IPs. We offer Residential, ISP, Unlimited Residential, Mobile, and Datacenter proxies for reliable data access across AI, web scraping, and automation workflows. Sign up using the promo code DRWEB to enjoy an exclusive 10% discount!

---

How to use

Inspect evidence and follow up

1. After a research finishes, click a [1]-style citation in the report, or click Inspect evidence and follow up.
2. In the evidence panel, review each finding's source, retrieval time, and excerpt. A new excerpt is only saved once it's matched against retrieved content; search summaries and page text are labeled separately. A matched excerpt does not by itself prove the finding — read it in context.
3. Enter what you want to verify, e.g. "Find the latest official pricing and confirm the eligibility terms," then click Follow up and update report. A follow-up searches at most two directions in one round, and updates only the Markdown blocks that cite that finding; tables and lists are updated as whole blocks.
4. On success, the new evidence and updated report are saved as a separate history entry, leaving the original report unchanged. On failure, cancellation, or when no matching excerpt is found, the original report is preserved.

This works in both Client Mode and Server Mode. Older history entries still load; findings without a saved excerpt are labeled accordingly. History is stored in your browser — export research you want to keep.

Live demo: https://deep-research.ataw.top

Deployment modes

• Client Mode: users enter their own API keys in the browser. This is the best fit for static deployments such as EdgeOne Pages or pnpm generate.
• Server Mode: API keys are configured as server-side environment variables, so users do not need to enter keys in the UI. This requires an SSR/Nitro runtime such as the Docker image; it is not available in purely static deployments.

Self hosted

Server Mode (Recommended)

Deploy with environment variables - users don't need to configure API keys. Use this mode when you can run the Nuxt server:

Using Docker with environment variables:

docker run -p 3000:3000 \
  -e NUXT_PUBLIC_SERVER_MODE=true \
  -e NUXT_AI_API_KEY=your-ai-api-key \
  -e NUXT_WEB_SEARCH_API_KEY=your-search-api-key \
  -e NUXT_PUBLIC_AI_PROVIDER=openai-compatible \
  -e NUXT_PUBLIC_AI_MODEL=gpt-4o-mini \
  -e NUXT_PUBLIC_WEB_SEARCH_PROVIDER=tavily \
  anotia/deep-research-web:latest

Using Docker with .env file:

# Copy .env.example and update it with your configuration
cp .env.example .env
docker run -p 3000:3000 --env-file .env anotia/deep-research-web:latest

Client Mode (Traditional)

Users configure their own API keys in the browser. Use this mode for static deployments:

One-click deploy with EdgeOne Pages:

Use pre-built Docker image:

docker run -p 3000:3000 --name deep-research-web -d anotia/deep-research-web:latest

Use self-built Docker image:

git clone https://github.com/AnotiaWang/deep-research-web-ui
cd deep-research-web-ui
docker build -t deep-research-web .
docker run -p 3000:3000 --name deep-research-web -d deep-research-web

Environment Variables

Server Mode Configuration

Public Configuration (Server Mode)

Provider values

Notes:

• When NUXT_ACCESS_PASSWORD is set, visitors must enter it before running research. Strongly recommended for publicly reachable server-mode deployments, since the server's API keys are otherwise usable by anyone.
• NUXT_WEB_SEARCH_API_KEY supports comma-separated keys for Tavily, Google PSE and Serply, for example key1,key2,key3.
• Google PSE requires both NUXT_WEB_SEARCH_API_KEY and NUXT_PUBLIC_GOOGLE_PSE_ID.
• Firecrawl self-hosted deployments can set NUXT_WEB_SEARCH_API_BASE.
• fastCRW (crw) is a Firecrawl-compatible web scraper (single binary; self-host or cloud). It defaults to the cloud base https://fastcrw.com/api and reads the key from NUXT_WEB_SEARCH_API_KEY (document as CRW_API_KEY); self-hosted deployments can set NUXT_WEB_SEARCH_API_BASE.
• You.com (youcom) reads its key from NUXT_WEB_SEARCH_API_KEY (optional, comma-separated keys supported for rotation). Without a key it uses the keyless endpoint with a limited daily quota; get a key at https://you.com/platform/api-keys.
• Serply (serply) returns Google web and news results through the Serply API and reads its key from NUXT_WEB_SEARCH_API_KEY (comma-separated keys supported for rotation). Time range, news intent, language and domain filters are applied natively; explicit publication-date windows are not.
• Ollama uses http://localhost:11434/v1 as the default API base. When running the app inside Docker, localhost refers to the container itself, so set NUXT_AI_API_BASE to a reachable host or Docker network address if Ollama runs outside the container.
• LiteLLM uses http://localhost:4000/v1 as the default API base. Its API key is optional when the proxy does not require authentication; set NUXT_AI_API_BASE when the proxy is not reachable at the default local address.
• Requesty uses https://router.requesty.ai/v1 as the default API base and expects model IDs in provider/model format, such as openai/gpt-4o.

Parallel Search MCP (Server Mode only)

Set these variables in your server deployment:

NUXT_PUBLIC_SERVER_MODE=true
NUXT_PUBLIC_WEB_SEARCH_PROVIDER=parallel

Keep your AI provider configured as usual. Parallel search needs no Parallel account or API key and does not use NUXT_WEB_SEARCH_API_KEY or NUXT_WEB_SEARCH_API_BASE. Tavily remains the default when no search provider is selected.

This provider uses the free, rate-limited Parallel Search MCP at https://search.parallel.ai/mcp. Searches send queries and objectives to Parallel; on-demand page reading sends source URLs. Related calls share a random research session ID. Requests include deep-research-web-ui/<app version> as their User-Agent so Parallel can measure aggregate project usage. See Parallel's Customer Terms and Privacy Policy.

Public Server Mode deployments share one outbound IP, so users share the anonymous free-tier rate limits. Reduce concurrency if requests are rate-limited.

Search excerpts retain their URLs, titles and publication dates. The app can also request full source pages when excerpts are insufficient. News, time, domain and language filters are unsupported and appear through the app's existing limitation notices. The result count is capped locally. Service errors are surfaced without switching providers.

Parallel is available only in Server Mode, including follow-ups. Static deployments and the Client Mode provider picker do not support it because browsers cannot reliably send the project User-Agent. Server requests honor the outbound proxy settings below.

Outbound proxy (Server Mode only)

Set NUXT_PROXY_URL to route server-side requests through a proxy:

NUXT_PROXY_URL=socks5h://user:[email protected]:7777
NUXT_NO_PROXY=localhost,127.0.0.1,::1

• Supported schemes: http, https, socks5, socks5h, socks.
• http(s) proxies cover everything: AI providers, Google PSE, you.com, Serply, Parallel, and the Tavily / Firecrawl / CRW SDKs.
• socks* proxies cover AI providers, Google PSE, you.com, Serply and Parallel; the Tavily / Firecrawl / CRW SDKs (axios-based) cannot speak SOCKS and will connect directly. Use the http:// endpoint of the same proxy (same credentials on most providers) for full coverage.
• NUXT_NO_PROXY accepts *, exact hosts and parent domains (example.com also matches api.example.com). Local AI gateways (Ollama, LiteLLM) and self-hosted scrapers stay direct via the default list.
• Standard HTTP_PROXY / HTTPS_PROXY / NO_PROXY env vars are honored as fallback when NUXT_PROXY_URL is unset.
• Proxy credentials are redacted in logs.
• Client Mode (browser) cannot use this: browsers have no SOCKS API. Configure a system / browser proxy instead.
• URL-encode special characters in credentials (% as %25, @ as %40, : as %3A).
• An invalid NUXT_PROXY_URL fails server startup (fail-fast) instead of silently going direct.

Using a proxy (e.g. Helodata) with self-hosted deployments

If your server needs geo-targeted or rotating egress IPs, run the deployment behind a standard HTTP/SOCKS5 proxy. Example with Helodata:

# Full coverage: AI + Google PSE + you.com + Tavily/Firecrawl SDKs
NUXT_PROXY_URL="http://YOUR_HELODATA_USERNAME:[email protected]:7777"
# Or SOCKS5: AI + Google PSE + you.com
# (Tavily/Firecrawl SDKs can't speak SOCKS and will connect directly)
NUXT_PROXY_URL="socks5h://YOUR_HELODATA_USERNAME:[email protected]:7777"
docker run -p 3000:3000 \
  -e NUXT_PUBLIC_SERVER_MODE=true \
  -e NUXT_AI_API_KEY=your-ai-api-key \
  -e NUXT_WEB_SEARCH_API_KEY=your-search-api-key \
  -e NUXT_PROXY_URL="$NUXT_PROXY_URL" \
  anotia/deep-research-web-ui:latest

Get credentials from your Helodata dashboard. See Helodata docs for region / city / sticky-session username format. For proxy behavior details, see Outbound proxy (Server Mode only).

---

Developing

Setup

Make sure to install dependencies:

pnpm install

Development Server

Start the development server on http://localhost:3000:

pnpm dev

Production

Build the application for production:

If you want to deploy a SSR application:

pnpm build

If you want to deploy a static, SSG application:

pnpm generate

Locally preview production build:

pnpm preview

Check out the deployment documentation for more information.

License

MIT

Star History

Acknowledgements

The project is also hosted on AtomGit, which is more accessible from mainland China. Thanks to AtomGit for the hosting.