Champion Council - AI Model Orchestrator

Local-first AI orchestration for VS Code and compatible editors. Run multi-model workflows, semantic memory, diagnostics, and community collaboration from one extension.

Core Features

• Multi-Slot Model Council — Plug HuggingFace models into council slots and run inference, debate, consensus, and chaining. Up to 32 slots, capsule-driven. Smart loader auto-detects model type (LLM, Embedding, Reranker, SEQ2SEQ, Vision, VLM, Classifier) from actual config.json metadata and routes to the correct loader. Cross-encoders are loaded via CrossEncoder() with sentence-pair scoring. Token budgets are read from each model's actual config (max_position_embeddings, generation_config) at plug time — no hardcoded limits. Rich metadata cards show model author, task, downloads, license, and size.
• 140+ MCP Tools — Full MCP/SSE tool surface for IDE agents and automation. Works with any MCP client (Claude Code, Cursor, Windsurf, etc.).
• Workflow Engine v2.2 — 10 node types: tool (universal MCP tool caller), agent (plugged model with granted tools loop), input, output, fan_out (parallel), http, if, set, merge, and web_search. Build DAGs with expression interpolation, conditional branching, agentic reasoning loops, and live execution tracing.
• Semantic Memory (FelixBag) — Local embedding store with search, catalog, induction, and export. Persistent across sessions.
• Activity Feed — Real-time tool call monitoring from all connected MCP clients. Click to expand full args, results, and timing. External agent activity detected via log polling across Windsurf, Cursor, VS Code, Kiro, and Antigravity.
• Community (Nostr) — Decentralized marketplace, live chat, encrypted DMs, NIP-57 zaps, and privacy controls.
• P2P Voice Rooms — Real-time voice communication via PeerJS/WebRTC with Nostr relay signaling (NIP-53). No server hosting required — fully peer-to-peer.
• Model Evaluation — Per-slot and council-level inference quality tracking with latency percentiles, consistency scoring, and best-combination analysis.
• Reputation Chains — Merkle-linked W3C Verifiable Credential history for tamper-evident publisher trust.
• Web3 Utilities — IPFS CID generation, DID:key derivation, Verifiable Credentials, Safety Attestations. Zero external APIs.
• IPFS Pinning — Pin content to Pinata or web3.storage directly from the extension.
• Glass Box Visualization — Optional Rerun integration for real-time inference pipeline visualization.

Voice Communications

Champion Council includes fully decentralized P2P voice rooms — no server hosting, no accounts, no third-party dependencies.

How It Works

1. Room Discovery — NIP-53 voice room events (kind 30312) published to Nostr relays
2. Peer Signaling — WebRTC signaling via Nostr ephemeral events (kind 25050)
3. Mic Capture — ffmpeg captures raw PCM audio from the OS mic in the extension host (Node.js)
4. Audio Bridge — PCM streams over a local WebSocket to the webview, where an AudioWorklet converts it to a MediaStream
5. Audio Transport — PeerJS sends the MediaStream over WebRTC peer connections with public STUN servers
6. Speaking Detection — Real-time audio level analysis via Web Audio API AnalyserNode

Voice Features

• Create & Join Rooms — Create voice rooms visible to all connected Nostr users, or join existing ones
• Real-Time Mic Feedback — Live level bar (green → yellow → red) shows your mic input in real-time
• Voice Settings — Mic sensitivity slider (0.5x–4.0x) and noise gate threshold (0–30) in the active room settings panel
• Speaking Indicators — See who's talking with per-participant audio level detection
• Live Chat — Text chat alongside voice in active rooms (kind 1311)
• Raise Hand — Signal to speak in moderated rooms
• Privacy Toggle — Enable/disable voice features entirely from the Privacy tab

Nostr Protocol Coverage

Quick Start

1. Install the extension.
2. Ensure Python 3.10+ is available on PATH.
3. Open Champion Council in the Activity Bar.
4. The MCP server auto-starts (default port 8765).
5. Open Control Panel to manage models, workflows, memory, and community.

Voice Quick Start

1. Go to the Community tab → Voice sub-tab
2. Click CREATE A ROOM and give it a name
3. Click MIC ON — grant microphone access when prompted
4. The level bar shows your mic input in real-time
5. Other users joining the same room will auto-connect via P2P
6. Tune sensitivity and noise gate via the gear icon in the active room view

Using with MCP Clients

The extension runs an MCP server on http://127.0.0.1:8765/sse. Any MCP-compatible client can connect:

• Claude Code — Add to .claude/settings.json or run Champion: Generate MCP Config for IDE
• Cursor / Windsurf — Add the SSE endpoint to your MCP configuration
• Custom clients — Connect via SSE transport to the endpoint above

All 140+ tools are available to any connected client. The Control Panel's activity feed shows tool calls from all sources.

Requirements

• Python 3.10+
• numpy (auto-installed by backend)
• torch (optional, for GPU inference)
• transformers (optional, for HuggingFace model loading)

Configuration

All settings live under champion.* in VS Code Settings.

Commands

• Champion: Open Control Panel
• Champion: Start MCP Server
• Champion: Stop MCP Server
• Champion: Generate MCP Config for IDE

What's New in 0.8.9

• RERANKER Dispatch Fix — CrossEncoder (reranker) models no longer crash dispatch tables. Fixed across _invoke_model, _council_deliberate, and all 6 hasattr-based MCP tools (pipe, compare, broadcast, debate, chain, all_slots), plus invoke_slot auto mode. Mixed councils with rerankers now work everywhere.
• Per-Councilor Fault Isolation — One bad model in the council can no longer crash the entire deliberation. Each councilor is wrapped in try/except with hash-based fallback.

See CHANGELOG.md for full release history.

Architecture

See COMMS_ARCHITECTURE.md for the full communications and identity protocol roadmap.

License

MIT