fiber-mcp

FIBER MCP Server

mcp-name: io.github.liveplex-cpu/fiber

Model Context Protocol (MCP) server for the FIBER AI Infrastructure Reliability Platform.

Expose 24 curated FIBER endpoints as MCP tools (+ 3 write tools behind a flag) so Claude Code / Cursor / Claude Desktop / VS Code Copilot can query optical telemetry, run Digital Twin what-if scenarios, fetch ML predictions, and generate postmortems — directly from developer IDEs.

Why MCP?

Until now, using FIBER meant opening the dashboard or calling the REST API. With MCP, your AI assistant just… knows.

User (in Claude Code): "leaf-010 Ethernet1/43 의 72시간 RUL 어떻게 돼?"

Claude  → calls tool: fiber_ml_predict_rul(switch_name="leaf-010", port_name="Ethernet1/43")
         → returns: {"rul_hours": 0.01, "urgency": "imminent", "val_mae_hours": 2.99}

Claude (to user): "LSTM 모델이 잔존 수명을 0.01시간으로 예측합니다 —
                 즉시 교체 필요 (urgency: imminent). 모델 검증 MAE는 2.99h입니다."

Tools (24 read + 3 write)

Category	Tools
Discovery	fiber_health, fiber_info, fiber_cluster_overview, fiber_list_ports, fiber_port_detail_history, fiber_list_alerts, fiber_top_risky_ports
ML	fiber_ml_rul_info, fiber_ml_predict_rul, fiber_ml_anomaly_scan
CPO	fiber_cpo_bank_status
Digital Twin	fiber_simulate_link_failure, fiber_digital_twin_monte_carlo, fiber_digital_twin_compare
Counterfactual	fiber_counterfactual_port, fiber_counterfactual_cluster
NCCL	fiber_nccl_summary, fiber_nccl_optical_correlation
Benchmark	fiber_run_benchmark_harness
Sustainability	fiber_sustainability, fiber_checkpoint_aware_windows
Diagnosis	fiber_list_diagnosis_reports, fiber_get_diagnosis_report, fiber_postmortem_timeline
Write (optional)	fiber_trigger_demo_scenario, fiber_ingest_nccl_events, fiber_create_rma

Safety model: all tools are read-only by default. Write tools only activate when FIBER_ENABLE_WRITE=1 env var is set. The server uses the same API key / RBAC as the underlying FIBER REST API, so giving an MCP client a viewer-role key is strictly safer than handing out admin credentials.

Install

Quick

./setup.sh

Manual

python3 -m venv ~/.fiber-mcp-venv
source ~/.fiber-mcp-venv/bin/activate
pip install -r requirements.txt

Configure

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "fiber": {
      "command": "/Users/you/.fiber-mcp-venv/bin/python3",
      "args": ["/path/to/hikari/sdks/mcp/fiber_mcp_server.py"],
      "env": {
        "FIBER_API_URL": "https://api.fiber.ac",
        "FIBER_API_KEY": "your-key-here",
        "FIBER_ENABLE_WRITE": "0"
      }
    }
  }
}

Restart Claude Desktop. In a new conversation ask: "What's FIBER's current cluster overview?" — Claude should call fiber_cluster_overview and respond with live data.

Claude Code (CLI)

Claude Code auto-discovers MCP servers configured in ~/.claude/mcp_servers.json:

{
  "fiber": {
    "command": "python3",
    "args": ["/path/to/fiber_mcp_server.py"],
    "env": {
      "FIBER_API_URL": "https://api.fiber.ac",
      "FIBER_API_KEY": "your-key"
    }
  }
}

Cursor

Cursor follows the same MCP spec. Add the server under Settings → MCP → Edit Config.

Test locally with MCP Inspector

npx -y @modelcontextprotocol/inspector python3 fiber_mcp_server.py

Opens a web UI where you can list tools and invoke them interactively.

Environment variables

Variable	Default	Purpose
FIBER_API_URL	https://api.fiber.ac	Base URL of your FIBER deployment
FIBER_API_KEY	fiber-dev-key-change-in-production	Bearer token (get via admin API)
FIBER_ENABLE_WRITE	0	Set to 1 to expose write tools
FIBER_HTTP_TIMEOUT	30	Per-request seconds

RBAC hints

• For AI assistants on an engineer's laptop, create a viewer role API key. All 22 read tools work; no accidental mutations.
• For an autonomous agent (e.g., oncall runbook runner), use operator + FIBER_ENABLE_WRITE=1.
• For demo environments, pair admin + FIBER_ENABLE_WRITE=1 to unlock fiber_trigger_demo_scenario.

User flow examples

Natural-language queries your AI assistant now handles:

• "How many CPO banks are in critical state?"
• "Show me the top 10 risky ports by risk score"
• "If we lose all 10 spine uplinks of leaf-010, how many GPUs go offline?"
• "Run the benchmark harness with seed 42 and summarize"
• "What's the checkpoint-aware maintenance window for the next 6 hours?"
• "Correlate NCCL bandwidth drops with optical anomalies over the last 15 minutes"
• "Generate a postmortem timeline for leaf-039 Ethernet1/43 covering the past 24h"

Architecture

 Claude Code / Cursor / Claude Desktop
              │
              │  MCP stdio (JSON-RPC)
              ▼
 ┌──────────────────────────┐
 │  fiber_mcp_server.py     │ ── Python stdio MCP server
 │                          │    (25 tools, httpx client)
 └────────────┬─────────────┘
              │  HTTPS Bearer auth
              ▼
 ┌──────────────────────────┐
 │  https://api.fiber.ac    │ ── FIBER FastAPI (~45 endpoints)
 │  or localhost:8000       │
 └──────────────────────────┘

Status

• Curated 27 tools (24 read + 3 write)
• stdio transport (SSE transport planned for remote/cloud use)
• Read-only by default; write tools guarded by env var
• Verified with @modelcontextprotocol/inspector

Roadmap

• Streaming SSE endpoint for cloud deployment
• OAuth device flow (replace long-lived API keys)
• GraphQL subscription tool (live port updates)

License

Same as FIBER — see repo root.