Parchi is a free, MIT-licensed AI browser copilot — a Chrome MV3 and Firefox 109+ extension written in TypeScript that puts chat-driven browser automation in a side panel, built for people who want to navigate, read, click, type, extract, and summarize pages under their own model configuration.
What it is
Parchi lives in the browser-extension ecosystem as an add-on for Chrome (Manifest V3) and Firefox 109 and later. A streaming chat panel sits beside the page and drives a set of browser tools: navigate, read, interact, tabs, and screenshots. The project is deliberately model-agnostic, so the same panel can point at OpenAI at https://api.openai.com/v1, an Anthropic-compatible endpoint, OpenRouter at https://openrouter.ai/api/v1, a local runtime such as Ollama or LM Studio at http://localhost:11434/v1, or any other OpenAI-compatible endpoint. The source is a TypeScript monorepo split into packages/extension for the runtime, UI, and tools, packages/backend for a Convex backend handling auth, billing, and proxying, and packages/shared for schemas, prompts, and runtime types, with scripts/ and tests/ alongside.
The concrete problem it solves is the gap between what a user reads in a tab and what an assistant can act on. Instead of copying text out of a page into a chat window, writing throwaway bookmarklets, or pasting ad-hoc scripts per site, the agent operates in the same tab context under explicit tool permissions and a domain allowlist. It also removes the need to accept whichever cloud copilot a vendor ships with the browser, because a user profile can be repointed at a different provider, or at a local model, without changing the extension.
Key capabilities
- Streaming chat paired with a tool execution timeline, so each model action is visible as it runs.
- Browser tools covering navigate, read, interact, tabs, and screenshots.
- Profile-based model and provider configuration with API key, endpoint or base URL, model ID, and optional custom headers.
- An orchestrator plus subagent flow for multi-step tasks.
- Session history with exports.
- Tool permissions and domain allowlist controls for restricting what the agent may touch.
- Separate build targets: npm run build produces the Chrome bundle in dist/, and npm run build:firefox produces the Firefox bundle in dist-firefox/.
Who uses it and how
- Chrome users build the extension, enable Developer mode, and load the unpacked dist/ directory from chrome://extensions; Firefox users load dist-firefox/ through about:debugging.
- Cost- or privacy-sensitive users point a profile at a local Ollama or LM Studio endpoint on http://localhost:11434/v1 so page content and prompts stay on their own machine.
- Automation-oriented users lean on the browser-automation and workflow topics, running repeated read, extract, and summarize loops from the side panel rather than by hand.
- Contributors work against the harnesses in tests/, running npm run test:unit, npm run test:integration, and npm run test:e2e, along with the orchestrator and performance harnesses.
Getting started
Build from source: git clone the repository, run npm install, then npm run build, and load the resulting dist/ folder as an unpacked extension; Firefox builds with npm run build:firefox and loads from about:debugging. The project also lists a homepage at https://parchi.ai.
How it compares
The facts provided name no paid products that Parchi replaces and no directly comparable tools, so it stands alone in this registry rather than being positioned against a named commercial alternative. Its distinguishing axis, as documented, is model-agnostic configuration inside a browser extension under an MIT licence, with no stated vendor lock-in to an endpoint.
When to use it — and when not to
Parchi suits users willing to build from source, supply an API key or run a local model, and accept the safety notice, since browser automation can violate a site's terms of service and prompt injection can trigger unsafe actions if untrusted prompts or pages are involved. Anyone who runs packages/backend takes on operating a Convex backend for auth, billing, and proxying. It is a poor fit for people who want a managed, zero-configuration copilot, and the current state shows a modest 546 stars, 54 forks, and 11 open issues, so it is best treated as a self-hosted tool for hands-on users rather than a polished consumer product.
