Take a look here https://github.com/brasa-interactive-studios/brasa-ai-core-engine

Local-first modular AI system for VSCodium/VSCode with server orchestration, web control panel, and integrated Claude Code CLI support.

Chat Interface: Same look & feel as GitHub Copilot — streams responses with markdown formatting, editor context awareness, and automatic fallback handling.

⚠️ IMPORTANT: The Claude Code CLI account currently has NO CREDITS. See TROUBLESHOOTING.md for solutions (use local Ollama or hybrid mode).

• Existing AI core modules (context, prompt, routing, generation, provider abstraction)
• New monorepo scaffolding with pnpm + turborepo
• New ai-server (Fastify + WebSocket) for centralized config and streaming updates
• New control-panel (React + Tailwind + Vite) for live visibility into shared/common config
• New vscode-extension lightweight bridge for editor sync + streamed chat over WebSocket
• New shared package for strongly typed cross-app contracts and defaults

apps/
  ai-server/
  control-panel/
  vscode-extension/          # scaffold placeholder

packages/
  shared/
  core/                      # transition placeholder
  sdk/                       # scaffold placeholder

src/                         # current core runtime (pre-extraction)

Responsibilities currently implemented:

• GET /health basic service health and websocket client count
• GET /api/config retrieve common orchestration config
• PATCH /api/config patch and persist config
• GET /api/repositories list repository registry
• POST /api/repositories create/update repository entries
• GET /ws/config websocket stream for config snapshots/updates + telemetry pulses
• GET /ws/extension websocket endpoint for VSCode extension messages (editor.sync, chat.request, chat.cancel)
• GET /api/prompt-traces prompt trace records for observability/prompt inspection

Persistence:

• Config is stored in apps/ai-server/.data/common-config.json

Web UI features currently implemented:

• Live connection status
• Local/cloud model overview
• Repository registry panel
• Routing diagnostics panel
• Live telemetry pulse panel
• Prompt trace panel (recent routes, model, latency, token estimates)
• Full JSON config inspection panel

Provides typed contracts used by both server and frontend:

• CommonConfig
• RepositoryWorkspace
• createDefaultConfig()
• mergeCommonConfig()

Current runtime remains in src/ and is already unit-tested.

Main modules:

• Context engine
• Prompt builder
• Ollama provider
• Model router
• Generation engine
• Diff applier
• Side chat controller

• TypeScript
• Node.js 20+
• Fastify + WebSocket
• React + Vite + Tailwind
• pnpm workspaces + Turborepo
• Ollama local inference (primary local model: qwen2.5-coder:14b)

• Node.js 20+ (via nvm recommended)
• Ollama (for local models, optional for cloud-only mode)
• Claude Code CLI (for cloud Claude route, optional for local-only mode)
• pnpm 9+ (Node package manager)

export NVM_DIR="$HOME/.nvm"
[ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"
nvm use 20

pnpm install

If you want to use Claude Code as a cloud backend:

⚠️ Current Status: Claude CLI is already installed and authenticated, but the account has NO CREDITS.

Choose one option:

A) Use Local Ollama Instead (Recommended) ✅

• Free, fast, completely local
• No API costs or internet needed
• Skip to step 4 and set providerMode to local-ollama

B) Add Claude Credits (if you want cloud Claude)

# Check current account:
claude auth status

# If loggedIn:true but subscriptionType:null, you need credits
# Visit https://console.anthropic.com and add payment method
# Then set providerMode to "copilot-managed" or "hybrid"

For detailed troubleshooting, see TROUBLESHOOTING.md

Local server + web panel:

./start-all.sh

This will:

• Start ai-server on http://127.0.0.1:8080/health
• Start control-panel on http://127.0.0.1:5173
• Auto-detect Claude CLI binary and export CLAUDE_CODE_BIN

To stop:

./stop-all.sh

The extension is located in apps/vscode-extension/ and provides:

• Real-time editor sync (file, cursor, diagnostics)
• Chat participant for integrated AI requests
• Sidebar commands for server control and Claude auth
• Automatic Claude auth recovery on errors

1. Build the extension:

pnpm --filter @apps/vscode-extension build

2. Install in VSCodium:

cp -r apps/vscode-extension ~/.config/VSCodium/extensions/ai-platform-extension
# or symlink for development:
ln -s $(pwd)/apps/vscode-extension ~/.config/VSCodium/extensions/ai-platform-extension

3. Enable extension in VSCodium:

Open VSCodium, go to Extensions sidebar → Install from VSIX → select apps/vscode-extension/dist/ or reload if symlinked.

# Package as VSIX
pnpm --filter @apps/vscode-extension build
# Then drag .vsix file into VSCode Extensions sidebar, or:
code --install-extension apps/vscode-extension/*.vsix

Open the AI Platform sidebar (left activity bar):

• Ask AI — send prompt via chat participant
• Connect Server — manually connect to WebSocket
• Start Server — run ./start-all.sh
• Stop Server — run ./stop-all.sh
• Claude Auth Login — open terminal for claude auth login --console
• Claude Auth Status — check current auth status
• Set Provider Mode — switch between copilot-managed, local-ollama, hybrid
• Cancel Active Request — abort current chat

Build all packages:

pnpm build:all

Type-check:

pnpm typecheck:all

Watch mode (local dev):

pnpm --filter @apps/ai-server dev
pnpm --filter @apps/control-panel dev
pnpm --filter @apps/vscode-extension dev

Run tests:

pnpm test:all

The platform now supports an inspectable control-plane baseline and is prepared for:

• Prompt trace persistence
• Token analytics dashboards
• Routing decision diagnostics
• Indexing/telemetry visibility
• Local/cloud orchestration transparency

• Core AI unit tests passing (21 tests)
• Server/control-panel scaffolding implemented
• Shared config contracts established
• VSCode extension and SDK packages scaffolded for next phase
• Claude Code CLI integration complete with hybrid routing
• VSCodium extension fully functional with sidebar commands
• Terminal-based auth flow with automatic error recovery

The system supports three provider routing modes:

1. copilot-managed — Use Claude Code CLI for all requests (default)

   • Requires: claude binary installed and authenticated
   • Fallback: Automatic local Ollama if Claude unavailable
   • Models: Configurable via CLAUDE_CODE_MODEL (default: sonnet)

2. local-ollama — Use local Ollama-only routing

   • Requires: Ollama service running
   • Models: qwen2.5-coder:14b, llama2, etc.
   • No cloud API keys needed

3. hybrid — Smart routing based on model capabilities

   • Attempts Claude Code first
   • Falls back to Ollama if Claude errors or unavailable
   • Best for development/testing with max coverage

VSCodium/VSCode (Extension)
         ↓
    WebSocket (WS)
         ↓
   ai-server (Fastify)
   ├→ ws/extension (chat, editor sync)
   ├→ ws/config (config changes + telemetry)
   ├→ api/config (REST config access)
   └→ Model Router (Ollama or Claude)
         ↓
    Local Ollama OR Claude Code CLI

Claude Code CLI path is resolved in this order:

1. CLAUDE_CODE_BIN — explicit path (set by start-all.sh)
2. command -v claude — system PATH lookup
3. Fallback candidates:
   • ~/.local/bin/claude (Linux pip install)
   • /opt/homebrew/bin/claude (macOS Homebrew)
   • /usr/local/bin/claude (manual install)

The start-all.sh script auto-detects and exports the path for child processes.

Conductor's chat interface provides identical user experience to GitHub Copilot while offering more flexibility:

Bottom Line: Same clean chat UI as Copilot, but you choose the backend (local Ollama, cloud Claude, or both with automatic fallback).

{
  "providerMode": "hybrid",
  "defaultLocalModel": "qwen2.5-coder:14b",
  "claudeCodeModel": "sonnet",
  "claudeCodeTimeout": 30000,
  "repositories": []
}

{
  "aiPlatform.serverUrl": "ws://127.0.0.1:8080/ws/extension",
  "aiPlatform.providerMode": "hybrid",
  "aiPlatform.claudeBinPath": "~/.local/bin/claude"
}

• CLAUDE_CODE_BIN — Path to Claude CLI executable
• CLAUDE_CODE_MODEL — Model name (sonnet, opus, haiku)
• CLAUDE_CODE_DEFAULT_MODEL — Fallback if not set
• CLAUDE_CODE_TIMEOUT_MS — Request timeout in milliseconds

Example:

export CLAUDE_CODE_BIN="/home/usuario/.local/bin/claude"
export CLAUDE_CODE_MODEL="opus"
export CLAUDE_CODE_TIMEOUT_MS="60000"

Symptom: Extension shows error "Claude: not logged in"

Solution:

1. Open the AI Platform sidebar in the extension
2. Click Claude Auth Login (opens terminal)
3. Run the suggested auth command:

   claude auth login --console

4. Copy the browser link and paste it into the terminal
5. Complete the auth flow, then retry your chat request

Alternative: Manual auth in system terminal:

claude auth login --console

Symptom: claude command works but returns: "Credit balance is too low"

Root Cause: Claude Code CLI account doesn't have active API credits. This happens when:

• Free trial has expired
• Subscription is inactive
• No payment method on file
• Credits were exhausted

Solutions:

1. Check your Claude account:

   claude auth status

   Look at subscriptionType — if it's null or free-expired, you need credits.

2. Add Claude subscription (Recommended if you want Claude):

   • Visit console.anthropic.com
   • Add a payment method
   • Enable API access
   • Revert providerMode to copilot-managed

3. Use Local Mode (Recommended for development):

   • Set providerMode to local-ollama or hybrid
   • Conductor will use Ollama (free, local) instead
   • No API costs, full privacy
   • Slightly slower but fully functional

Current Status in Conductor:

# Check what's available:
$ CLAUDE_CODE_MODEL=sonnet echo "test" | claude -p
Error: Credit balance is too low  ❌

# But local models work:
$ curl http://127.0.0.1:11434/api/generate -d '{"model":"qwen2.5-coder:14b", "prompt":"test"}' 
[Streams response]  ✅

Recommended: Use hybrid mode (default) — Conductor will try Claude first, automatically fall back to Ollama if credits are exhausted.

Symptom: ./start-all.sh fails or port 8080 already in use

Solution:

# Kill existing process on port 8080
lsof -i :8080 | grep LISTEN | awk '{print $2}' | xargs kill -9

# Or explicitly stop:
./stop-all.sh

# Then restart:
./start-all.sh

Symptom: "Failed to connect to server" message in sidebar

Solution:

1. Verify server is running:

   curl -s http://127.0.0.1:8080/health | jq .

2. Check extension settings (Ctrl+, → search "AI Platform"):

   • serverUrl should be ws://127.0.0.1:8080/ws/extension
   • serverUrl must be ws:// or wss:// (WebSocket)

3. Click Connect Server in sidebar to retry

4. Check browser console (F12 in extension dev tools) for WebSocket errors

Symptom: Chat takes >30 seconds or times out

Solution:

1. Increase timeout:

   export CLAUDE_CODE_TIMEOUT_MS="60000"
   ./stop-all.sh && ./start-all.sh

2. Check model:

   export CLAUDE_CODE_MODEL="haiku"  # faster, less capable

3. Check Claude CLI directly:

   echo "test prompt" | claude -p --model sonnet

Symptom: "Ollama service unreachable" when using local-ollama mode

Solution:

1. Start Ollama:

   ollama serve

2. In another terminal, pull a model:

   ollama pull qwen2.5-coder:14b

3. Switch to local mode in extension:

   • Sidebar → Set Provider Mode → select local-ollama

4. Verify connection:

   curl -s http://127.0.0.1:11434/api/tags | jq .