Local-first web interface for exploring your X/Twitter bookmarks. Built on the fieldtheory CLI.

Read-only · Zero cloud · Dark mode

• Search, filter, and explore thousands of bookmarks with full-text search
• Ask natural language questions and get interactive dashboards back
• See how your interests evolve over time, who you bookmark, and what repos you're collecting

Requires: Node.js v20+ and fieldtheory CLI v1.2.0+ with synced bookmarks.

git clone https://github.com/Gitmaxd/field-theory-ui.git
cd field-theory-ui
npm install
./start.sh

Open http://localhost:3939. The start script builds, launches, and prints a QR code for mobile access on your local network.

Chat is an AI-powered interface for querying your bookmarks. It operates in two modes with increasing capability, backed by a LangChain ReAct agent that autonomously selects tools to answer your questions.

Works out of the box with no API keys. Uses Claude Code CLI, Codex CLI, or built-in pattern matching (first available) to translate natural language into API calls against your bookmark collection.

Adds a LangChain ReAct agent with tool calling and interactive dashboards -- charts, tables, KPI cards, and follow-up buttons rendered as live React components via OpenUI. Enable it by adding an API key:

cp .env.example .env
# Set ANTHROPIC_API_KEY or OPENAI_API_KEY in .env

A PRO toggle appears in the Chat input bar when a key is detected. The agent auto-detects your provider and selects the appropriate model (anthropic:claude-sonnet-4-6 or openai:gpt-5.4-mini, overridable via CHAT_MODEL). When using Anthropic, the system prompt is cached across agent invocations to reduce latency and token cost on multi-step queries. Pro mode markdown responses include Copy and Download actions for extracting LLM-generated analysis.

The Pro mode agent has access to the following tools and autonomously decides which to call, in what order, chaining up to 8 tool invocations per query:

When both tools are available, the agent queries your bookmarks first for local context, then searches the web for current information, then synthesizes both sources in its response.

Example prompts showing tool usage:

Database only: "Build a dashboard comparing my top 5 categories by volume over the last 6 months" -- the agent writes aggregation queries via execute_sql and renders an interactive OpenUI dashboard with charts and KPI cards.

Web search only: "What are the latest developments in AI agent frameworks?" -- the agent calls TavilySearch for current web results and responds with a markdown summary of its findings.

Combined (local + web): "Research the topics I've been bookmarking about AI agents and tell me what I'm missing" -- the agent queries your bookmarks via execute_sql to map your AI agent content, then uses TavilySearch for current developments, then synthesizes the gaps between what you've saved and what's happening now.

Combined (verification): "I bookmarked several posts about Rust -- is it still gaining traction?" -- the agent pulls your Rust bookmark data for context, searches the web for current adoption metrics, and compares the two.

Pro mode streams responses in real time with step-level progress indicators ("Thinking...", "Querying database...", "Generating response..."). Text responses stream with progressive markdown rendering and a blinking cursor. Dashboard responses render progressively during streaming as the OpenUI Renderer assembles interactive components in real time. Both Anthropic and OpenAI token formats are handled natively.

The input auto-expands up to 7 lines (Shift+Enter inserts a newline). The Surprise Me button fetches a random bookmark from your collection.

npm run dev              # Vite dev server with HMR
npm run server           # API server only (port 3939)
npm run build            # Production build
npm run typecheck        # tsc --noEmit
npm run lint             # ESLint
npm test                 # Vitest
npm run generate:prompt  # Regenerate OpenUI component reference for Chat Pro

npm update -g fieldtheory
ft sync --gaps

Built by @GitMaxd -- github.com/GitMaxd

MIT