How yannicksiewe/skyops-coder works
skyops-coder is a self-hosted engineering AI platform: team tools send chat, code, image, or agent requests through compatible model endpoints, while separate workflows provide fine-tuning and repository automation. Its documented runtime also includes gateway accounting and tracing, monitoring, and encrypted backups. The README supports the major deployment relationships; sampled source confirms agent orchestration, shared LLM/GitHub helpers, diff parsing, MCP requests, and tool-call parsing. Other operational links are shown only where documented, not inferred from configuration names.
Main entry points: vllm-chat.service Chat model · gateway/compose.yml LLM gateway · ai-task.yml PR and deploy workflows · monitoring/compose.yml Prometheus and Grafana
Components 25
Model serving
- Chat model
serving/vllm-chat.service - Autocomplete model
serving/vllm-autocomplete.service - Vision model
serving/vllm-vision.service - Tool-call parser
serving/coder_tool_parser.py
Team interfaces
- Open WebUI
serving/webui_config.py - IDE and SDK clients
clients/continue-config.yaml - Aider client
clients/aider.env - MCP tools
clients/mcp/skyops_mcp.py
Repository agents
- Code review agent
agents/review.py - Architecture review agent
agents/arch_review.py - Issue-to-PR agent
agents/issue_to_pr.sh - Agent LLM and GitHub
agents/common.py - Diff parser
agents/diffparse.py
Model training
- LoRA fine-tuning
training/train.py - Adapter inference
training/infer.py
Platform operations
- LLM gateway
ops/gateway/compose.yml - Secrets guardrail
ops/gateway/guardrails/secrets_guardrail.py - Self-hosted runner
ops/runner/install_runner.sh - PR and deploy workflows
.github/workflows/ai-task.yml - Request tracing
ops/langfuse/compose.yml - Prometheus and Grafana
ops/monitoring/compose.yml - Encrypted backups
ops/backup/backup.sh
Other components
- Engineering team
- Browser user
- GitHub
Connections 27
- Engineering team → IDE and SDK clients: uses
README.md - Engineering team → Aider client: uses
README.md - Engineering team → MCP tools: invokes tools
README.md - Browser user → Open WebUI: uses
README.md - IDE and SDK clients → Chat model: requests chat
README.md - IDE and SDK clients → Autocomplete model: requests completion
README.md - Aider client → Chat model: requests chat
README.md - Open WebUI → Chat model: connects to
README.md - Open WebUI → Autocomplete model: connects to
README.md - MCP tools → LLM gateway: sends API requests
clients/mcp/skyops_mcp.py - LLM gateway → Chat model: routes requests
README.md - LLM gateway → Vision model: routes requests
README.md - Chat model → Tool-call parser: extracts tool calls
serving/coder_tool_parser.py - PR and deploy workflows → Self-hosted runner: runs on
README.md - PR and deploy workflows → Code review agent: runs review
README.md - PR and deploy workflows → Architecture review agent: runs review
README.md - Code review agent → Diff parser: parses diffs
agents/review.py - Code review agent → Agent LLM and GitHub: uses LLM and GitHub
agents/review.py - Architecture review agent → Diff parser: parses diffs
agents/arch_review.py - Architecture review agent → Agent LLM and GitHub: uses LLM and GitHub
agents/arch_review.py - Issue-to-PR agent → GitHub: creates PR
- LoRA fine-tuning → Adapter inference: produces adapter
- LLM gateway → Secrets guardrail: applies redaction
README.md - LLM gateway → Request tracing: records traces
README.md - Prometheus and Grafana → Chat model: monitors services
README.md - Prometheus and Grafana → Autocomplete model: monitors services
README.md - Prometheus and Grafana → Vision model: monitors services
README.md