AvaxOps

AI-native operations layer for Avalanche validators and L1s.

🔗 Live site: avax-ops.vercel.app

Ask your AI agent: "Which of my validators need attention this week?" — and get a real answer.

Components

• mcp-server/ — Open-source MCP server exposing Avalanche ops tools (validator health, expirations, balances, diagnostics) to Claude, Cursor, Claude Code, or any MCP-compatible agent. MIT licensed.
• service/ (planned, M2) — Monitoring backend: polls P-Chain and node health, evaluates alert thresholds (uptime → 90% ACP-267 limit, expiring validations, low L1 balances, version lag).
• bot/ (planned, M2) — Telegram bot for node registration, real-time alerts, and weekly fleet digests.

Status

v0.1 — MCP server with 5 working tools, verified against live Fuji and mainnet P-Chain APIs. Built for the Avalanche Team1 Builder Grants program; see docs/grant-application.md for the roadmap.

MCP tools

Tool	What it does
ping	Server + P-Chain connectivity check (network, chain height).
list_validators	Fleet table: stake (AVAX), uptime, connectivity, delegators, start/end, days remaining.
expiration_report	Validators whose validation period ends within N days (default 30).
validator_health	Per-validator report incl. the ACP-267 90% uptime check; optionally reads your node's own Info/Health APIs (version, bootstrap status, health checks).
balance_check	P-Chain address balances (total/unlocked/locked) and L1 validator continuous-fee balances with runway estimate.

All tools are read-only and zero-config for public data. No keys, no signing, no accounts.

Install

Requires Python 3.11+ and uv.

git clone https://github.com/EMPangolin/avax-ops.git
cd avax-ops
uv venv
uv pip install -e mcp-server

Use it from Claude Code

claude mcp add avax-ops -e AVAX_NETWORK=mainnet -- \
  uv --directory /path/to/avax-ops run avax-ops-mcp

Then just ask:

"Show me all my validators expiring in the next 30 days." "Is NodeID-7Xhw2mDx... meeting the ACP-267 uptime requirement?" "How much AVAX is left on my L1 validator's fee balance?"

Works the same in Claude Desktop, Cursor, or any MCP client (stdio transport, command avax-ops-mcp).

Configuration

Env var	Default	Meaning
AVAX_NETWORK	mainnet	mainnet or fuji (testnet).

Example output

validator_health against a live Fuji validator:

{
  "nodeID": "NodeID-CRVEYUqnqNvCSBnWGQ8c1Fmg2qHZe581E",
  "found": true,
  "network": "fuji",
  "connected": true,
  "uptime_pct": 99.9958,
  "uptime_meets_acp267_90pct": true,
  "stake_AVAX": 1.0,
  "end": "2026-08-18 05:30 UTC",
  "days_remaining": 40.3
}

Development

cd mcp-server
uv pip install --group pyproject.toml:dev   # pytest, pytest-asyncio, ruff
pytest tests/ -q                            # unit tests (RPC layer mocked)
ruff check src tests && ruff format --check src tests

API field names (e.g. stake arriving as weight in nAVAX, not the legacy stakeAmount) are verified against the live P-Chain API; test fixtures mirror real captured responses.

Roadmap

• M1 — MCP server v0.1, read-only ops tools (this release)
• M2 — Monitoring service + Telegram alerts (uptime, expirations, L1 balances)
• M3 — AI diagnostics (diagnose_node), top-up estimator, L1 config helper
• M4 — Web dashboard, docs, demo

License

MIT