Skip to content

Releases: rish-e/Autopilot

v1.2.0 — Browser Stability

Choose a tag to compare

@rish-e rish-e released this 25 Mar 17:43

Autopilot v1.2.0 — Browser Stability

Configures Playwright MCP with Chromium stability flags and a persistent browser profile, significantly reducing browser deaths during long sessions.

What changed

Playwright stability config (new)

Chromium flags that prevent the most common causes of browser death:

  • Disabled background timer throttling
  • Disabled renderer backgrounding
  • Disabled occluded window backgrounding
  • Disabled IPC flooding protection
  • Disabled hang monitor
  • Disabled back-forward cache
  • Navigation timeout increased from 60s to 120s

Persistent browser profile (new)

  • Profile directory at ~/MCPs/autopilot/browser-profile/
  • Cookies, localStorage, and sessions survive across Claude Code restarts
  • Login to a service once, stay logged in across sessions

Installer auto-configuration (updated)

The installer now auto-configures the Playwright MCP in .claude.json with stability flags and persistent profile path.

Performance impact

These flags reduce CPU usage — they tell Chrome to stop running background monitoring checks.

v1.1.0 — Browser Recovery and MCP Safety

Choose a tag to compare

@rish-e rish-e released this 25 Mar 17:43

Autopilot v1.1.0 — Browser Recovery & MCP Safety

Fixes a critical issue where the agent would attempt to terminate the browser process when it died, which permanently disconnected the MCP server.

What changed

Browser Recovery Protocol (new)

When the Playwright browser dies mid-session, the agent now:

  1. Never attempts to terminate MCP server processes
  2. Immediately checks if CLI tools can handle the task instead
  3. Falls back to CLI with stored Keychain credentials
  4. Only asks user to restart their session if browser is truly required

Guardian rules for MCP process protection (new)

  • 6 new custom rules block process termination commands targeting browser automation and MCP server processes
  • These are hard blocks (exit code 2) — the agent cannot bypass them

Updated agent definition

  • New Core Principle #6: "NEVER TOUCH MCP PROCESSES"
  • New "Browser Recovery Protocol" section
  • New error handling case for dead browser/MCP servers

Test suite expanded

  • 4 new guardian tests (45 total, up from 41)

v1.0.0 — Initial Release

Choose a tag to compare

@rish-e rish-e released this 25 Mar 17:43

Autopilot v1.0.0 — Initial Release

The first self-expanding autonomous coding agent with hard safety rails.

What's included

  • Autopilot agent definition — fully autonomous Claude Code agent
  • Guardian hook — PreToolUse safety layer with tested block patterns
  • macOS Keychain integration — encrypted credential storage
  • 5 service registries — Vercel, Supabase, GitHub, Cloudflare, Razorpay
  • Decision framework — 5-level system for autonomous action classification
  • MCP auto-discovery — whitelist-based trust system for MCP servers
  • Self-expansion — agent researches unknown services and teaches itself
  • Smart permissions — speed of --dangerously-skip-permissions with safety
  • One-command installer — curl pipe or clone + run

Requirements

  • macOS (uses macOS Keychain)
  • Claude Code installed
  • Homebrew + Node.js (installer handles these)