Future

BrowserShell v1.0 is a working shell with 105 commands. The architecture is deliberately extensible. Here’s what’s next.

Package system

Goal: Extend the shell without forking the extension — Homebrew for browser commands.

pkg search weather
pkg install weather
pkg update
weather "San Francisco"

Design principles

Packages are signed bundles of commands, VFS providers, and optional page scripts
Installed to /packages/<name>/ in the virtual filesystem
Registry hosted on GitHub (community contributions via PR)
Sandboxed execution — packages declare required permissions upfront
pkg uninstall removes cleanly

Why it matters

Today, adding a command requires a TypeScript PR. The package system lets the community ship integrations for Jira, Linear, Gmail, HN, Reddit, internal tools — without waiting for core releases.

Site hooks

Goal: Websites expose domain-specific commands when you arrive.

When BrowserShell loads a page, it checks for a site manifest:

{
  "name": "github",
  "commands": {
    "pr": { "usage": "pr list", "handler": "bs://github/pr-list" },
    "issue": { "usage": "issue <n>", "handler": "bs://github/issue" }
  }
}

Delivered via:

<link rel="browser-shell" href="/browser-shell.json"> in page HTML
Or well-known path: /.well-known/browser-shell.json

What this enables

Site	Commands
GitHub	`pr list`, `issue 42`, `repo clone`
Gmail	`mail unread`, `mail search`
Your SaaS	`deploy status`, `logs tail`
Internal wiki	`search docs`

Sites opt in. No hook = generic commands still work (links, read, go).

Security

Site manifests can only register commands scoped to their origin
User confirmation on first activation per site
Manifests are cached and inspectable via siteinfo --hooks

AI-native shell

Goal: When AI can interact with the browser, the shell is the API.

Today’s ai summarize uses Chrome’s on-device AI. The next step is bidirectional:

# You ask
ai "close all YouTube tabs and summarize my last 5 HN articles"

# The agent plans and executes
→ tabs | grep youtube     # 3 tabs found
→ tab close 2; tab close 4; tab close 7
→ history search hackernews | head -5
→ ai summarize (5 pages)

Why a shell, not a chatbot?

Chatbot	Shell
Opaque actions	Every step is a visible, editable command
Can’t compose	Pipes, aliases, scripts
No audit trail	`/audit` logs every action
Black box	You can interrupt, modify, re-run

The shell becomes the integration layer between humans and agents — agents propose command sequences, humans approve or edit them, the executor runs them deterministically.

Building blocks (already in v0.1)

Structured command output (--json flags)
Audit log (/audit)
Transcript (/transcript)
Watch mode for monitoring
Composable pipes

Horizon

Agent mode: natural language → command plan → user confirms → execute
MCP integration: BrowserShell as a tool server for external agents
Multi-step macros: record and replay command sequences
Context-aware prompts: AI reads /current/content.txt and suggests next commands

Chrome Web Store

Planned listing with:

Privacy policy ✅
Permissions rationale ✅
Security policy ✅
Store screenshots and demo video (coming)

Get involved

The roadmap is public and open source. Pick an area:

Commands — add core commands via PR
Packages — help design the pkg spec (discussion welcome on GitHub)
Site hooks — prototype browser-shell.json on your site
AI — experiment with agent → command translation

Contributing on GitHub →