About

Built by people who ship in browsers.

Linda started as a server-side experiment called lichat. We ripped out the server, ripped out the dependencies, and rebuilt it as a browser-first agent SDK. This is the result.

Why Linda exists

Most "in-page AI" today is a chatbot widget in a sidebar — a thin UI wrapped around a stateless chat completion. That works for a Q&A bot, but it falls apart the moment you want the AI to do something: understand your form, fill its fields, validate them against rules, hand off to a specialist agent at the right moment, ingest a dropped PDF without uploading it, gate on browser capabilities, talk MCP to your CRM.

Linda is what we wished existed for that work. The bet is simple: modern LLMs are trained on filesystem-shaped workflows (Claude Code, agentic repo edits, MCP servers). So give them a filesystem — but make it the live page. Reads from /page/dom.html are the DOM snapshot. Writes to /page/form/fields/email.json are how the agent fills a field. Everything else falls out of that.

What's special

  • Browser-first. The agent loop runs in the browser. Files stay in the browser. The model is reached via fetch + SSE. No axios, no multer, no sharp.
  • Composable primitives. VFS, tool loop, hooks, skills, multi-agent, MCP, parsers. Each is independently swappable.
  • No vendor lock. Five LLM providers out of the box, plus OpenAI-compatible baseUrl override. Bring your own model.
  • Small. ~35 KB core, ~34 KB UMD. Parsers are opt-in.
  • MIT. Forever. No CLA, no rug-pull.

The team

Linda is built by Dipankar Sarkar and the Neul Labs team. Neul Labs builds small, open, opinionated developer infrastructure.

Other Neul Labs projects

  • closegate — open-source policy gate for finance AI agents.
  • BondFoundry — AI on the buy-side fixed-income desk; FINOS AIGF v2.0 reference implementation.

How to get involved

Ship an agent-driven flow this afternoon.

Install Linda, paste a config, and your form turns into an agent that fills its own inputs.