How rwilliamspbg-ops/ghostlink works
Ghostlink is a self-hosted distributed inference fabric for heterogeneous LAN hardware, with an OpenAI-compatible API and a Studio UI. The main workflow is a user submitting a chat request through the UI or API; the backend runs inference locally or uses LAN peers, then returns the result. Separate product branches cover cluster visibility and management, MCP tools, and developer SDKs. The README documents the product flow and capabilities; sampled source and the SOURCE INDEX clarify selected UI, gateway, inference, and core relationships, but many backend routes remain unsampled.
Main entry points: main.rs Inference API · App.tsx Studio application · registry.rs MCP tool registry · api.ts UI API client
Components 24
User interfaces
- Studio application
ghostlink_gui_modern/src/App.tsx - Chat workspace
ghostlink_gui_modern/src/components/ChatTab.tsx - Model controls
ghostlink_gui_modern/src/components/ModelsTab.tsx - Cluster dashboard
ghostlink_gui_modern/src/components/WorkersTab.tsx - UI API client
ghostlink_gui_modern/src/api.ts
Inference serving
- Control-plane gateway
control-plane/main.go - Inference API
crates/ghost-link/src/main.rs - Native engine adapter
crates/ghost-link/src/native_engine.rs - Ollama engine
crates/ghost-link/src/ollama.rs - vLLM engine
crates/ghost-link/src/vllm.rs
Cluster fabric
- Peer discovery
crates/ghostlink-core/src/discovery.rs - Placement planning
crates/ghostlink-core/src/planning.rs - Cluster coordination
crates/ghostlink-core/src/cluster.rs - RPC model sharding
crates/ghost-link/src/rpc_cluster.rs - Cluster protocol
crates/ghostlink-core/src/protocol.rs
Tools and integrations
- MCP tool registry
crates/ghost-link/src/mcp/registry.rs - Durable memory tool
crates/mcp-memory/src/main.rs - RAG tool server
crates/mcp-rag/src/main.rs - Vision tool server
crates/mcp-vision/src/main.rs
Client access
- JavaScript SDK
sdks/js/src/client.ts - Python SDK
sdks/python/ghostlink_client/client.py
Other components
- User
- llama.cpp server
- API consumer
Connections 28
- User → Studio application: uses
README.md - Studio application → Chat workspace: renders
ghostlink_gui_modern/src/App.tsx - Studio application → Model controls: renders
ghostlink_gui_modern/src/App.tsx - Studio application → Cluster dashboard: renders
ghostlink_gui_modern/src/App.tsx - Chat workspace → UI API client: requests
ghostlink_gui_modern/src/components/ChatTab.tsx - Model controls → UI API client: requests
- Cluster dashboard → UI API client: requests
ghostlink_gui_modern/src/components/WorkersTab.tsx - UI API client → Control-plane gateway: HTTP requests
README.md - Control-plane gateway → Inference API: proxies requests
control-plane/main.go - API consumer → Inference API: submits requests
README.md - Inference API → Native engine adapter: dispatches inference
- Native engine adapter → llama.cpp server: sends generation requests
crates/ghost-link/src/native_engine.rs - Native engine adapter → RPC model sharding: uses cluster adapter
crates/ghost-link/src/native_engine.rs - Inference API → Ollama engine: supports engine
- Inference API → vLLM engine: supports engine
- Peer discovery → Cluster coordination: forms LAN cluster
README.md - Placement planning → Cluster coordination: plans placement
crates/ghostlink-core/src/planning.rs - Cluster coordination → Cluster protocol: uses messages
crates/ghostlink-core/src/cluster.rs - Placement planning → Cluster protocol: uses protocol types
crates/ghostlink-core/src/planning.rs - RPC model sharding → Cluster coordination: distributes inference
README.md - Studio application → MCP tool registry: configures tools
- MCP tool registry → Durable memory tool: connects tool server
- MCP tool registry → RAG tool server: connects tool server
- MCP tool registry → Vision tool server: connects tool server
- API consumer → JavaScript SDK: uses SDK
sdks/js/src/client.ts - API consumer → Python SDK: uses SDK
- JavaScript SDK → Inference API: calls API
sdks/js/src/client.ts - Python SDK → Inference API: calls API