›_ Claude Code Issues
Claude Code Issues / #27938

[FEATURE] Open-Source Toolkit: Automated Loop Driver, Council Automation, MCP Browser Bridge (Bun Workaround), and More for Claude Code CLI

Resolved 💬 3 comments Opened Feb 23, 2026 by intellegix Closed Apr 13, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code CLI is powerful, but out of the box it lacks several capabilities that power users and teams need for production-grade autonomous workflows:

  1. No autonomous loop driver — Running Claude Code in long-running autonomous loops requires manual session management, budget tracking, stagnation detection, and model-aware timeout scaling. Users currently have to babysit sessions or build their own orchestration.
  1. No multi-model council/synthesis — There's no built-in way to query multiple AI models (GPT, Claude, Gemini) simultaneously and synthesize results, which is critical for plan validation and architectural decisions.
  1. Chrome extension Bun compatibility issues on Windows — The official "Claude in Chrome" extension relies on Bun for its native host, and Bun has a known panic on Windows when creating named pipes (Bun Canary v1.3.9 assertion failure). This has been reported in multiple duplicate issues (#24034, #22416, #23823, #21935, #24248, #24280, #27670) and remains unresolved, preventing Windows users from connecting Chrome to Claude Code entirely.
  1. No reusable slash command library — Teams rebuild the same research, planning, code review, and deployment workflows from scratch in every project.
  1. No portfolio governance — No built-in system for managing complexity budgets, project tiers, or phase restrictions across multiple projects.

Proposed Solution

I've built and open-sourced a modular toolkit that addresses all of the above problems. It's available now under MIT license:

Repository: https://github.com/intellegix/intellegix-code-agent-toolkit**

What it includes:

1. Automated Loop Driver (automated-loop/)

  • Runs Claude Code CLI in autonomous loops via claude -p --stream-json
  • NDJSON stream parsing, session continuity (--resume), and budget enforcement
  • Model-aware scaling (Opus: 2x timeout, 25-turn cap) with automatic fallback (Opus -> Sonnet)
  • Stagnation detection with two-strike system and exponential backoff
  • 194 pytest tests

2. MCP Browser Bridge — Fixes the Bun/Windows Chrome Extension Problem (mcp-servers/browser-bridge/)

  • A WebSocket-based bridge between Claude Code CLI and Chrome that completely bypasses Bun
  • Uses a standard Node.js MCP server + a Chrome MV3 extension communicating over ws://127.0.0.1:8765
  • This directly solves the Bun named pipe panic on Windows documented in #24034 (and duplicates #22416, #23823, #21935, #24248, #24280, #27670) by avoiding Bun's native host architecture entirely
  • Drop-in replacement: load the extension unpacked, start the Node server, configure via mcp.json

3. Council Automation (council-automation/)

  • Multi-model queries via Perplexity (GPT, Claude, Gemini) with Opus synthesis
  • Playwright-based browser automation using your existing Perplexity Pro subscription ($0/query)
  • Session cookie management with 24h TTL and auto-refresh

4. Custom Slash Commands (commands/)

  • 15+ reusable commands: /research-perplexity, /smart-plan, /council-refine, /fix-issue, /implement, /review, /handoff, and more
  • Uses $ARGUMENTS variable for flexible invocation

5. Portfolio Governance (portfolio/)

  • Tier-based project management (T1 Production through T4 Archive)
  • Complexity budgets, phase restrictions, and testing requirements per tier

Installation:

git clone https://github.com/intellegix/intellegix-code-agent-toolkit.git ~/.claude

Full setup docs, architecture diagrams, and configuration examples are in the README.

Alternative Solutions

For the Chrome extension Bun issue specifically, the alternatives I considered before building the MCP Browser Bridge:

  • Wait for upstream Bun fix — The Bun named pipe panic has been reported since at least #21935 and remains open across 7+ duplicate issues. Waiting is not viable for Windows users who need browser integration now.
  • Downgrade Bun — Claude Code ships with a specific Bun version; downgrading introduces other compatibility risks.
  • Use Node.js native messaging — Would still require modifying Claude Code's internals.

The MCP Browser Bridge takes a different architectural approach: a Node.js WebSocket server + Chrome MV3 extension that sidesteps Bun entirely while providing the same browser automation capabilities.

For the other components (loop driver, council, commands), I couldn't find existing open-source solutions that integrate cleanly with Claude Code's CLI and CLAUDE.md conventions.

Priority

High - Significant impact on productivity

Feature Category

Developer tools/SDK

Use Case Example

Scenario: Windows user needs browser automation with Claude Code

  1. User installs Claude Code on Windows 11 and wants to use the Chrome extension for browser-integrated workflows
  2. Running claude.exe --chrome-native-host crashes immediately with a Bun assertion failure on named pipe creation (see #24034)
  3. User clones the toolkit: git clone https://github.com/intellegix/intellegix-code-agent-toolkit.git ~/.claude
  4. Loads the MCP Browser Bridge Chrome extension (load unpacked from ~/.claude/mcp-servers/browser-bridge/extension/)
  5. Starts the Node.js MCP server: cd ~/.claude/mcp-servers/browser-bridge && npm start
  6. Configures mcp.json with the browser-bridge server entry
  7. Browser automation now works on Windows — no Bun, no named pipes, just WebSocket over localhost

Scenario: Autonomous multi-session project execution

  1. Developer sets up CLAUDE.md with a multi-phase project plan
  2. Runs python loop_driver.py --project /path/to/project --max-iterations 50 --max-cost 25.00
  3. The loop driver autonomously executes Claude Code CLI with budget enforcement, stagnation detection, and session continuity
  4. If Opus times out consecutively, it auto-falls back to Sonnet
  5. Developer reviews .workflow/state.json and trace.jsonl between runs to audit progress
  6. Updates CLAUDE.md to steer the next loop run — the human stays in control while the loop handles execution

Additional Context

  • Repository: https://github.com/intellegix/intellegix-code-agent-toolkit
  • License: MIT
  • Languages: Python 57.2%, JavaScript 41.3%
  • Tests: 194 pytest tests for the loop driver
  • Related Bun issue: #24034 (and 7+ duplicates — this is a widespread Windows problem)
  • Architecture: Modular — each component can be used independently
  • No API keys needed for Perplexity features — uses browser session cookies from your existing subscription
  • Not affiliated with Anthropic — community-built, trademark notice included in NOTICE file
  • Contributions and feedback welcome

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗