Lower token payload
Typical output is around 2-5KB instead of ~1MB screenshots, reducing model overhead and latency.
Text-grid browser for AI agents
See the web as text, not screenshots.
TextWeb executes full JavaScript in Chromium, then converts pages into compact, spatially meaningful text grids your LLM can reason about directly.
npm install -g textweb
Why teams choose TextWeb
Preserved layout
Element positions and visual structure are retained in a deterministic text grid, unlike raw HTML or accessibility trees.
Actionable references
Interactive elements are annotated with refs like [9], so agents can reliably click, type, and scroll.
From screenshot blob to readable grid
Screenshot + vision model
~1 MB
Large binary payload, expensive vision inference, harder deterministic interaction mapping.
TextWeb output
~2-5 KB
[0] Hacker News [1] new | [2] past
1. [9] Show HN: TextWeb
142 points | [10] comments
[13:______________] [14 Search]
Quick start
textweb https://news.ycombinator.com
textweb --interactive https://github.com
textweb --json https://example.com
Integrations
MCP server
Use textweb-mcp with Claude Desktop, Cursor, Windsurf, Cline, and other MCP-compatible clients.
Function calling
Drop in tools/tool_definitions.json with OpenAI or Anthropic tool-calling workflows.
LangChain + CrewAI
Use included adapters in tools/langchain.py and tools/crewai.py.
HTTP + Node.js API
Run textweb --serve 3000 for REST endpoints or import AgentBrowser directly in Node.