ethpandaops MCP Server Usage Guide

Query Ethereum network data through the ethpandaops MCP server. Execute Python code in sandboxed containers with access to ClickHouse blockchain data, Prometheus metrics, Loki logs, and Dora explorer APIs.

Workflow

1. Discover – Read MCP resources to find available datasources and schemas
2. Find patterns – Use search tools to find query examples and runbooks
3. Execute – Run Python with execute_python using the ethpandaops library

Quick Reference

Discovering Data Sources

Read these resources to understand what data is available:

Finding Query Patterns

Search for example queries:

search_examples(query="block arrival time")
search_examples(query="attestation participation", category="attestations")

Search for investigation runbooks:

search_runbooks(query="network not finalizing")
search_runbooks(query="slow queries", tag="performance")

The ethpandaops Python Library

ClickHouse – Blockchain Data

from ethpandaops import clickhouse

# List available clusters
clusters = clickhouse.list_datasources()
# Returns: [{"name": "xatu", "database": "default"}, {"name": "xatu-cbt", ...}]

# Query data (returns pandas DataFrame)
df = clickhouse.query("xatu-cbt", """
    SELECT
        slot,
        avg(seen_slot_start_diff) as avg_arrival_ms
    FROM mainnet.fct_block_first_seen_by_node
    WHERE slot_start_date_time >= now() - INTERVAL 1 HOUR
    GROUP BY slot
    ORDER BY slot DESC
""")

# Parameterized queries
df = clickhouse.query("xatu", "SELECT * FROM blocks WHERE slot > {slot}", {"slot": 1000})

Cluster selection:

• xatu-cbt – Pre-aggregated tables (faster, use for metrics)
• xatu – Raw event data (use for detailed analysis)

Required filters:

• ALWAYS filter on partition key: slot_start_date_time >= now() - INTERVAL X HOUR
• Filter by network: meta_network_name = 'mainnet' or use schema like mainnet.table_name

Prometheus – Infrastructure Metrics

from ethpandaops import prometheus

# List instances
instances = prometheus.list_datasources()

# Instant query
result = prometheus.query("ethpandaops", "up")

# Range query
result = prometheus.query_range(
    "ethpandaops",
    "rate(http_requests_total[5m])",
    start="now-1h",
    end="now",
    step="1m"
)

Time formats: RFC3339 or relative (now, now-1h, now-30m)

Loki – Log Data

Always discover labels first. Before querying logs, fetch the available labels and their values so you can add the right filters. Unfiltered Loki queries are slow and may time out — label filters narrow the search at the storage level and are essential for efficient log retrieval.

from ethpandaops import loki

# Step 1: List instances
instances = loki.list_datasources()

# Step 2: Fetch all available labels
labels = loki.get_labels("ethpandaops")
print(labels)
# Example: ['app', 'cluster', 'ethereum_cl', 'ethereum_el', 'ethereum_network',
#           'instance', 'namespace', 'node', 'testnet', 'validator_client', ...]

# Step 3: Get values for a specific label to build your filter
networks = loki.get_label_values("ethpandaops", "testnet")
print(networks)  # e.g. ['fusaka-devnet-3', 'hoodi', 'sepolia', ...]

cl_clients = loki.get_label_values("ethpandaops", "ethereum_cl")
print(cl_clients)  # e.g. ['lighthouse', 'prysm', 'teku', 'nimbus', 'lodestar', 'grandine']

# Step 4: Query logs with label filters
logs = loki.query(
    "ethpandaops",
    '{testnet="hoodi", ethereum_cl="lighthouse"} |= "error"',
    start="now-1h",
    limit=100
)

Key labels for Ethereum log queries:

• testnet — network/devnet name (e.g. hoodi, fusaka-devnet-3)
• ethereum_cl — consensus layer client (e.g. lighthouse, prysm, teku)
• ethereum_el — execution layer client (e.g. geth, nethermind, besu)
• ethereum_network — Ethereum network name
• instance — specific node instance
• validator_client — validator client name

Log level formats vary by client. When filtering logs by severity, be aware that Ethereum clients format log levels differently:

• Keywords: CRIT, ERR, ERROR, WARN, INFO, DEBUG
• Structured fields: level=error, "level":"error", "severity":"ERROR"
• Shorthand: E, W, C

Start with |~ "(?i)(CRIT|ERR)" as a default filter. If it returns no results, fetch a few unfiltered log lines to identify the client’s format, then adapt the regex (e.g. |~ "level=(error|fatal)").

Dora – Beacon Chain Explorer

Discovering all Dora API endpoints:

Before using Dora, discover the full set of available API endpoints by fetching the Swagger documentation. The swagger page is always at <dora-url>/api/swagger/index.html.

1. First, get the Dora base URL for the network:

from ethpandaops import dora
base_url = dora.get_base_url("mainnet")
print(f"Swagger docs: {base_url}/api/swagger/index.html")

2. Then use WebFetch to read the swagger page at {base_url}/api/swagger/index.html to discover all supported API endpoints for that Dora instance. This is important because different Dora deployments may support different endpoints.

3. Use the discovered endpoints to make targeted API calls via the Python dora module or direct HTTP requests.

Common API usage:

from ethpandaops import dora

# Get network health
overview = dora.get_network_overview("mainnet")
print(f"Current epoch: {overview['current_epoch']}")
print(f"Active validators: {overview['active_validator_count']}")

# Check finality
epochs_behind = overview['current_epoch'] - overview.get('finalized_epoch', 0)
if epochs_behind > 2:
    print(f"Warning: {epochs_behind} epochs behind finality")

# Generate explorer links
link = dora.link_validator("mainnet", "12345")
link = dora.link_slot("mainnet", "9000000")
link = dora.link_epoch("mainnet", 280000)

Direct HTTP calls for endpoints not in the Python module:

from ethpandaops import dora
import httpx

base_url = dora.get_base_url("mainnet")
# Call any endpoint discovered from swagger
with httpx.Client(timeout=30) as client:
    resp = client.get(f"{base_url}/api/v1/<endpoint>")
    data = resp.json()

Storage – Upload Outputs

from ethpandaops import storage

# Save visualization
import matplotlib.pyplot as plt
plt.savefig("/workspace/chart.png")

# Upload for public URL
url = storage.upload("/workspace/chart.png")
print(f"Chart URL: {url}")

# List uploaded files
files = storage.list_files()

Session Management

Critical: Each execute_python call runs in a fresh Python process. Variables do NOT persist.

Files persist: Save to /workspace/ to share data between calls.

Reuse sessions: Pass session_id from tool responses for faster startup and workspace persistence.

Multi-Step Analysis Pattern

# Call 1: Query and save
from ethpandaops import clickhouse
df = clickhouse.query("xatu-cbt", "SELECT ...")
df.to_parquet("/workspace/data.parquet")

# Call 2: Load and visualize (pass session_id from Call 1)
import pandas as pd
import matplotlib.pyplot as plt
from ethpandaops import storage

df = pd.read_parquet("/workspace/data.parquet")
plt.figure(figsize=(12, 6))
plt.plot(df["slot"], df["value"])
plt.savefig("/workspace/chart.png")
url = storage.upload("/workspace/chart.png")
print(f"Chart: {url}")

Session Tools

manage_session(operation="list")     # View active sessions
manage_session(operation="create")   # Pre-create a session
manage_session(operation="destroy", session_id="...")  # Free a session

Error Handling

ClickHouse errors include actionable suggestions:

• Missing date filter → “Add slot_start_date_time >= now() - INTERVAL X HOUR“
• Wrong cluster → “Use xatu-cbt for aggregated metrics”
• Query timeout → Break into smaller time windows

Default execution timeout is 60s, max 600s. For large analyses:

• Use search_examples to find optimized patterns
• Break work into smaller time windows
• Save intermediate results to /workspace/

Notes

• Always filter ClickHouse queries on partition keys (slot_start_date_time)
• Use xatu-cbt for pre-aggregated metrics, xatu for raw event data
• Check python://ethpandaops resource for complete API documentation
• Use search_examples before writing complex queries from scratch
• Use search_runbooks to find runbooks for common workflows
• Upload visualizations with storage.upload() for shareable URLs
• NEVER just copy/paste/recite base64 of images. You MUST save the image to the workspace and upload it to give it back to the user.