Manage Pi-hole v6: DNS blocking, domains, clients, query analysis, DHCP, and multi-instance sync.
A production-grade MCP server for Pi-hole v6.
82 tools (84 with a second instance) | 9 prompts | 5 resources | Multi-instance + sync | Read-only mode + toolsets | Single Go binary
Gives AI assistants full control over your Pi-hole instance: DNS blocking, domain management, query analysis, statistics, network devices, DHCP, and system administration. Compatible with the Pi-hole v6 REST API.
PIHOLE_READ_ONLY=true exposes only the tools that cannot change anything, enforced on every call, not just the tool list a client happens to cache. PIHOLE_TOOLSETS narrows further to just the families you need, so a client asking "what can I use" gets a smaller, cheaper answer.io.github.hexamatic/pihole-mcp, installable by name from any client that supports registry install.Most MCP clients use the same configuration format. Add this to your client's config:
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
Then install the binary via one of the methods below.
The whole point is natural language, not tool names. A few examples of what to actually type:
| You ask | What runs |
|---|---|
| "Is Pi-hole blocking right now, and for how long?" | pihole_dns_get_blocking |
| "Pause blocking for 10 minutes, I need to test something." | pihole_dns_set_blocking |
| "Why is reddit.com blocked? Check every list it could be on." | pihole_search_domains |
| "What are the top blocked domains today, and does anything look like a false positive?" | pihole_stats_recent_blocked, the review_top_blocked prompt |
| "Point nas.home at 192.168.1.50." | pihole_local_dns_add |
| "Which of my blocklists haven't updated in 30+ days?" | pihole_lists_list |
| "Show me devices on the network Pi-hole doesn't recognise." | pihole_network_devices, the audit_network prompt |
| "Compare my two Pi-holes and tell me what's different." | pihole_instance_diff |
| "Back up my configuration before I change anything." | pihole_teleporter_export |
| "Which upstream resolver is slowest, and is it worth switching?" | pihole_stats_upstreams, the upstream_health prompt |
Set PIHOLE_READ_ONLY=true first if you want to ask questions without any risk of something being
changed. See Scoping the tool surface.
pihole-mcp is listed in the official MCP Registry as:
io.github.hexamatic/pihole-mcp
Clients that support registry install can add it by that name and will prompt for
PIHOLE_URL and PIHOLE_PASSWORD. The listing points at the ghcr.io image, so the
client needs a working Docker.
brew install hexamatic/tap/pihole-mcp
Installs on both macOS and Linux (Homebrew on Linux). On macOS the cask clears the quarantine attribute during install, so the binary runs without a Gatekeeper prompt.
scoop bucket add hexamatic https://github.com/hexamatic/scoop-bucket
scoop install pihole-mcp
go install github.com/hexamatic/pihole-mcp/cmd/pihole-mcp@latest
docker pull ghcr.io/hexamatic/pihole-mcp:latest
.deb and .rpm packages for Debian-based (Ubuntu, Raspberry Pi OS) and RPM-based (Fedora, RHEL) distributions are available on the Releases page.
# Debian / Ubuntu / Raspberry Pi OS
sudo dpkg -i pihole-mcp_X.Y.Z_linux_amd64.deb
# Fedora / RHEL / CentOS
sudo rpm -i pihole-mcp_X.Y.Z_linux_amd64.rpm
Pre-built binaries for Linux, macOS, and Windows (amd64 and arm64) are available on the Releases page.
Releases are checksummed, signed with keyless cosign, and ship SPDX SBOMs and SLSA build provenance. See SECURITY.md for the verification commands.
Once installed and configured, run the server directly with -check. It loads the configuration, contacts every Pi-hole you have configured, and reports each one by name:
PIHOLE_URL=http://192.168.1.2 PIHOLE_PASSWORD=your-password pihole-mcp -check
pihole-mcp 0.9.0
PASS primary (http://192.168.1.2) core v6.1.4, FTL v6.7
All 1 instance(s) reachable and authenticated.
It exits non-zero if any instance fails, and prints what to change. Worth running before you wire the server into a client: most clients hide the server's output, so a bad URL or password looks the same as a server that never started.
| Variable | Required | Default | Description |
|---|---|---|---|
PIHOLE_URL | Yes | - | Pi-hole base URL (e.g. http://192.168.1.2) |
PIHOLE_PASSWORD | Yes | - | Admin password or application password |
PIHOLE_READ_ONLY | No | false | Expose only tools that cannot change Pi-hole. See Scoping the tool surface. |
PIHOLE_TOOLSETS | No | all | Comma-separated toolset names to expose, e.g. dashboard,domains. Unset or all exposes everything. |
PIHOLE_REQUEST_TIMEOUT | No | 30s | HTTP request timeout |
PIHOLE_MAX_RETRIES | No | 3 | Retries after a failed Pi-hole API call. 0 disables. |
PIHOLE_RETRY_MAX_DELAY | No | 8s | Upper bound on a single backoff wait. |
PIHOLE_HTTP_AUTH_TOKEN | No | - | Shared bearer token required on every HTTP/SSE request. Minimum 16 characters. Unset means no authentication. |
PIHOLE_HTTP_AUTH_TOKEN_FILE | No | - | Path to a file holding the bearer token, so it stays out of ps. Mutually exclusive with PIHOLE_HTTP_AUTH_TOKEN. |
PIHOLE_RATE_LIMIT | No | 120 | Per-session requests-per-minute cap on the HTTP/SSE transports, under a per-address ceiling of four times that. 0 disables. |
PIHOLE_ALLOWED_ORIGINS | No | localhost,127.0.0.1,[::1] | Comma-separated Origin/Host allowlist for HTTP/SSE transports. The literal * disables enforcement (unsafe). Not authentication. |
PIHOLE_TRUSTED_PROXIES | No | - | Comma-separated IPs or CIDR blocks whose X-Forwarded-For the rate limiter believes. Unset means the header is ignored. |
PIHOLE_TLS_SKIP_VERIFY | No | false | Disable TLS certificate verification for Pi-hole connections. Only for instances serving self-signed certificates. A trusted certificate is preferable where possible. |
TZ | No | System timezone (UTC in Docker) | IANA timezone for rendered timestamps (e.g. Australia/Adelaide). Timezone data is embedded in the binary, so this works in the Docker image out of the box. |
OTEL_EXPORTER_OTLP_ENDPOINT | No | - | OpenTelemetry collector endpoint. Setting it enables tracing; ignored in slim builds. |
Application passwords are recommended for automation: they bypass TOTP 2FA and can be revoked independently.
PIHOLE_HTTP_AUTH_TOKEN, PIHOLE_HTTP_AUTH_TOKEN_FILE, PIHOLE_RATE_LIMIT, PIHOLE_ALLOWED_ORIGINS and PIHOLE_TRUSTED_PROXIES only apply to the http and sse transports; stdio is a single-process, single-user channel by definition and isn't gated. See Security before exposing either transport beyond loopback.
By default the server offers every tool, including 34 that change Pi-hole and 27 that are destructive. Two variables narrow that.
PIHOLE_READ_ONLY=true exposes only tools that cannot change anything. Write tools are absent from
tools/list and rejected if called anyway, so a client working from a cached listing cannot
reach one either.
PIHOLE_READ_ONLY=true # 49 tools instead of 82
PIHOLE_TOOLSETS selects which families of tools exist at all, which is worth doing on its own
because tools/list is a cost paid in every conversation:
PIHOLE_TOOLSETS=dashboard,domains,lists # blocklist curation, and nothing else
PIHOLE_TOOLSETS=all # the default, stated explicitly
The two intersect, and read-only always wins. PIHOLE_READ_ONLY=true with
PIHOLE_TOOLSETS=dns,stats gives the 14 tools that are in those toolsets and read-only.
The toolset names, what each contains and how many tools it holds are in
docs/TOOLS.md, generated from the same table the server filters with. An
unknown name fails at startup and lists every valid one; run pihole-mcp -check to see it, along
with how many tools your configuration actually exposes:
scope: read-only=true, toolsets=dns,stats 14 of 82 tools exposed
A published toolset name is never removed, renamed or narrowed. See CONTRIBUTING.md for the full promise.
To manage more than one Pi-hole, configure numbered instances instead of PIHOLE_URL/PIHOLE_PASSWORD:
| Variable | Required | Description |
|---|---|---|
PIHOLE_1_URL, PIHOLE_2_URL, … | Yes | Base URL of each instance (contiguous from 1) |
PIHOLE_1_PASSWORD, PIHOLE_2_PASSWORD, … | Yes | Password for the matching instance |
PIHOLE_1_NAME, PIHOLE_2_NAME, … | No | Friendly name (default instance-1, instance-2, …) |
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_1_URL": "http://192.168.1.2",
"PIHOLE_1_PASSWORD": "primary-password",
"PIHOLE_1_NAME": "downstairs",
"PIHOLE_2_URL": "http://192.168.1.3",
"PIHOLE_2_PASSWORD": "secondary-password",
"PIHOLE_2_NAME": "upstairs"
}
}
}
}
Every tool then accepts an optional instance argument, and every result is labelled with the instance it came from. Omit the argument to target the first instance; pass a name to target a specific one; pass instance=all on a read-only tool (e.g. pihole_padd, pihole_stats_summary) to query every instance concurrently and get back a single structured aggregate (per-instance results plus a success/failure summary, so one slow or unreachable instance no longer fails the whole call). State-changing tools require a single named instance. PIHOLE_URL and PIHOLE_1_URL are mutually exclusive.
When you run more than one Pi-hole, two extra tools appear for keeping them aligned:
pihole_instance_diff: compare two instances and see exactly what differs across adlists/allowlists, allow/deny rules (exact and regex), groups, clients, local DNS A/AAAA records, and CNAME records. It is read-only and writes nothing.pihole_instance_sync: push a source instance's configuration onto a target. It is deliberately cautious:
source of truth and the target; only the target is ever written to.confirm_token by default; nothing changes until you re-run with mode=apply and that token. If the configuration drifts between planning and applying, the token no longer matches and the apply is refused.prune=true.snapshot=false).Example: preview what the upstairs Pi-hole is missing relative to downstairs, then apply it.
pihole_instance_diff { "source": "downstairs", "target": "upstairs" }
pihole_instance_sync { "source": "downstairs", "target": "upstairs" } → returns a plan + confirm_token
pihole_instance_sync { "source": "downstairs", "target": "upstairs",
"mode": "apply", "confirm_token": "<token from the plan>" }
The Quick Start config above works for most clients. Expand the section below for client-specific instructions.
Add to your Claude Desktop configuration file:
| OS | Path |
|---|---|
| macOS | ~/Library/Application Support/Claude/claude_desktop_config.json |
| Windows | %APPDATA%\Claude\claude_desktop_config.json |
| Linux | ~/.config/Claude/claude_desktop_config.json |
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
Restart Claude Desktop after saving.
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
claude mcp add pihole \
-e PIHOLE_URL=http://192.168.1.2 \
-e PIHOLE_PASSWORD=your-password \
-- pihole-mcp
Verify with:
claude mcp list
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
Add to .vscode/mcp.json in your workspace:
{
"servers": {
"pihole": {
"type": "stdio",
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
Or add via the command palette: MCP: Add Server.
Note: VS Code uses
"servers"as the top-level key (not"mcpServers"), and requires"type": "stdio".
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
Add to ~/.cursor/mcp.json:
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
Add to ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
Open Cline settings > MCP Servers > Configure, and add:
{
"mcpServers": {
"pihole": {
"command": "pihole-mcp",
"env": {
"PIHOLE_URL": "http://192.168.1.2",
"PIHOLE_PASSWORD": "your-password"
}
}
}
}
If the server does not appear, run pihole-mcp -check in a terminal with the same environment variables to see whether it can reach your Pi-hole.
For clients that support Docker-based MCP servers:
{
"mcpServers": {
"pihole": {
"command": "docker",
"args": ["run", "-i", "--rm",
"-e", "PIHOLE_URL=http://192.168.1.2",
"-e", "PIHOLE_PASSWORD=your-password",
"-e", "TZ=Australia/Adelaide",
"-v", "/host/path/for/backups:/backups",
"ghcr.io/hexamatic/pihole-mcp:latest"]
}
}
}
Useful when you don't have Go installed or want to run the server on a remote host.
The -v mount is there for pihole_teleporter_export. Without it, a backup is written inside the container's own filesystem and is gone as soon as --rm removes the container on exit. Pass output_path pointing inside the mount (e.g. /backups/pihole.zip) to keep it on the host.
To check the container can reach your Pi-hole, run the same image with -check. Remember that localhost inside a container is the container itself, which is the usual cause of a failure here:
docker run --rm \
-e PIHOLE_URL=http://192.168.1.2 \
-e PIHOLE_PASSWORD=your-password \
ghcr.io/hexamatic/pihole-mcp:latest -check
A single Pi-hole exposes 82 tools. Configuring more than one adds pihole_instance_diff and pihole_instance_sync, for 84. Those two are registered only when there is a second instance to compare against, so a single-Pi-hole setup isn't shown tools it cannot use.
To expose fewer, see Scoping the tool surface.
The tables below are a summary; the full generated reference with every parameter is in docs/TOOLS.md.
| Tool | Description |
|---|---|
pihole_padd | One-call snapshot: queries, blocking, top domain/client, cache, versions, host health |
| Tool | Description |
|---|---|
pihole_dns_get_blocking | Get current DNS blocking status and timer |
pihole_dns_set_blocking | Enable/disable blocking with optional timer |
| Tool | Description |
|---|---|
pihole_local_dns_list | List local A and AAAA records (dns.hosts) |
pihole_local_dns_add | Point a hostname at an IP address; the record type follows the address family |
pihole_local_dns_delete | Remove a local A or AAAA record by IP address and hostname |
pihole_local_cname_list | List local CNAME records (dns.cnameRecords) |
pihole_local_cname_add | Point an alias at another name, with an optional TTL |
pihole_local_cname_delete | Remove a local CNAME record by its alias |
| Tool | Description |
|---|---|
pihole_stats_summary | Queries, blocking rate, clients, gravity size |
pihole_stats_top_domains | Top queried or blocked domains |
pihole_stats_top_clients | Most active clients by query count |
pihole_stats_upstreams | Upstream DNS server performance |
pihole_stats_query_types | Query type distribution (A, AAAA, MX, etc.) |
pihole_stats_recent_blocked | Recently blocked domains |
pihole_stats_database | Long-term database statistics |
pihole_stats_database_top_domains/top_clients/upstreams/query_types | Long-term database equivalents of the four tools above, for an arbitrary date range |
| Tool | Description |
|---|---|
pihole_domains_list | List allow/deny domains |
pihole_domains_add | Add domains (bulk supported) |
pihole_domains_update | Update domain entry |
pihole_domains_delete | Remove a domain |
pihole_domains_batch_delete | Remove multiple domains |
| Tool | Description |
|---|---|
pihole_groups_list/add/update/delete/batch_delete | Manage groups |
pihole_clients_list/suggestions/add/update/delete/batch_delete | Manage clients |
pihole_lists_list/add/update/delete/batch_delete | Manage blocklists/allowlists |
| Tool | Description |
|---|---|
pihole_queries_search | Search queries with 12 filters + cursor pagination |
pihole_queries_suggestions | Available filter values |
| Tool | Description |
|---|---|
pihole_info_system | Host, CPU, memory, disk, load, temperature |
pihole_info_version | Pi-hole component versions |
pihole_info_database | Database size and query count |
pihole_info_messages | FTL diagnostic messages |
pihole_info_dismiss_message | Dismiss a diagnostic message by ID |
pihole_info_client | The requesting client's own IP and connection info |
pihole_info_ftl | FTL process info: PID, privacy level, client/domain counts |
pihole_info_metrics | Live DNS and DHCP operational metrics (cache, inserts, evictions, leases) |
pihole_info_sensors | Hardware temperature sensors |
pihole_search_domains | Cross-list domain search |
pihole_auth_sessions/revoke_session | List and revoke active API sessions |
pihole_config_get/set | Read/modify Pi-hole configuration |
pihole_config_get_value/add_value/remove_value | Granular dotted-path config access |
pihole_config_properties | List read-only config keys (Pi-hole v6.6.1+) |
| Tool | Description |
|---|---|
pihole_action_gravity_update | Re-download blocklists |
pihole_action_restart_dns | Restart FTL DNS resolver |
pihole_action_flush_logs | Flush the query log |
pihole_action_flush_network | Flush the network table (Pi-hole v6.3+) |
pihole_network_devices/gateway/info/routes/interfaces | Network device and interface discovery |
pihole_network_delete_device | Remove a stale network device record |
pihole_dhcp_leases/delete_lease | DHCP lease management |
pihole_logs_dns/ftl/webserver | Log retrieval |
pihole_teleporter_export/import | Configuration backup and restore |
pihole_history_graph/clients | Activity history (in-memory) |
pihole_history_database/database_clients | Activity history (long-term database, arbitrary date range) |
| Tool | Description |
|---|---|
pihole_instance_diff | Compare configuration between two instances |
pihole_instance_sync | Reconcile a target instance towards a source (dry-run plan, then confirmed apply) |
Some tools accept optional parameters for controlling output:
detail (minimal | normal | full): controls response depth. Default: normal. Use minimal for one-line summaries, full for complete API data. Available on 15 tools.format (text | csv): output format for tabular data. Default: text. CSV saves roughly 8–23% tokens over the default text rendering, depending on the tool and row count. Available on 22 tools: pihole_domains_list, pihole_groups_list, pihole_lists_list, pihole_clients_list, pihole_queries_search, pihole_network_devices, pihole_dhcp_leases, pihole_config_properties, pihole_local_dns_list, pihole_local_cname_list, pihole_stats_top_domains, pihole_stats_top_clients, pihole_stats_upstreams, pihole_stats_query_types, pihole_stats_recent_blocked, pihole_stats_database_top_domains, pihole_stats_database_top_clients, pihole_stats_database_upstreams, pihole_history_graph, pihole_history_clients, pihole_history_database, and pihole_history_database_clients.Pre-built multi-step workflows for common tasks:
| Prompt | Description |
|---|---|
diagnose_slow_dns | Analyse upstream performance and identify bottlenecks |
investigate_domain | Check why a domain is blocked/allowed across all lists |
review_top_blocked | Identify false positives in top blocked domains |
audit_network | Discover unknown devices and unconfigured clients |
optimise_blocklists | Suggest list consolidation and cleanup |
daily_report | Comprehensive daily Pi-hole health summary |
security_audit | Review active sessions and auth config for unauthorised access |
weekly_trends | Compare DNS statistics week over week |
upstream_health | Deep performance analysis of upstream resolvers |
Read-only context an MCP client can pull in without calling a tool:
| URI | Description |
|---|---|
pihole://status | Blocking status, version, health |
pihole://summary | Query statistics |
pihole://clients/{client} | Configuration and groups for one client |
pihole://domains/{type}/{kind} | Domains on a list, e.g. deny/exact |
pihole://lists/{address} | Details of one blocklist or allowlist |
With more than one Pi-hole configured, each instance is also addressable directly: pihole://instances lists them, and pihole://<instance>/status and pihole://<instance>/summary read a named one. The unprefixed URIs above always read the first-declared instance.
By default, pihole-mcp uses stdio (standard for MCP). HTTP and SSE transports are also available:
# Default stdio (for Claude Desktop, Cursor, etc.)
pihole-mcp
# HTTP transport (for web-based MCP clients)
pihole-mcp -transport http -address localhost:8080
# SSE transport (deprecated, see below)
pihole-mcp -transport sse -address localhost:8080
Both examples bind loopback, so only this machine can reach them. Before changing that, read Security below: neither transport authenticates anything until you configure a token.
pihole-mcp implements MCP protocol revision 2025-11-25 and negotiates down for older clients (2025-06-18, 2025-03-26 and 2024-11-05 are all accepted).
SSE is deprecated. The HTTP+SSE transport was superseded by Streamable HTTP in the 2025-03-26 revision, and formally reclassified as Deprecated under the specification's feature lifecycle policy (SEP-2596), meaning it stays part of the spec, fully supported, for a minimum of twelve months before removal could even be proposed.
-transport sseis kept for older clients and still receives security fixes, but new deployments should use-transport http.
The http and sse transports have no authentication until you configure one. stdio is
unaffected: it is a single-process, single-user channel by definition.
The default listen address is localhost:8080, so out of the box only this machine can reach the
server. Change the bind address without setting a token and anyone who can reach the port can
control your Pi-hole: disable blocking, rewrite your blocklists, read every DNS query on the
network.
Bearer authentication (recommended whenever the bind address is not loopback). Set a shared
token and every request must carry it in an Authorization: Bearer header. Requests without it
are answered 401 before anything else runs. Tokens must be at least 16 characters.
export PIHOLE_HTTP_AUTH_TOKEN="$(openssl rand -base64 32)"
# A bind beyond loopback needs the allowlist extended too, or clients reach a
# 403 from the Host check before the token is ever looked at.
export PIHOLE_ALLOWED_ORIGINS="localhost,127.0.0.1,[::1],192.168.1.10"
pihole-mcp -transport http -address 0.0.0.0:8080
An environment variable set on a command line is visible in ps to every user on the host, so
the token can be read from a file instead. This is also the shape container platforms mount
secrets in:
export PIHOLE_HTTP_AUTH_TOKEN_FILE=/run/secrets/pihole-mcp-token
Set one or the other, not both. The token is compared in constant time, so a wrong token tells an attacker nothing about how much of it was right. Rejected requests are charged to a separate failed-authentication budget of roughly twenty attempts per address and then one a second, so the token cannot be guessed at line rate, a flood of wrong tokens cannot starve a client at the same address that does hold it, and the server logs the throttling once rather than once per attempt.
Binding a non-loopback address with no token logs an error at startup and carries on. It does not refuse to start, because putting a reverse proxy that authenticates in front of it is a legitimate deployment.
Origin and Host validation. Both headers must resolve to a host in PIHOLE_ALLOWED_ORIGINS
(default loopback only). Mismatches return 403. Missing Origin is allowed, because non-browser
MCP clients do not send one.
This is not authentication. It is DNS-rebinding protection, and it only protects browsers: a
page a victim visits cannot make their browser attach an Origin the allowlist accepts. Both
headers are chosen by the caller, so any client that is not a browser sets them to whatever the
allowlist wants. Use PIHOLE_HTTP_AUTH_TOKEN for access control.
export PIHOLE_ALLOWED_ORIGINS="localhost,127.0.0.1,[::1],pihole-mcp.lan"
The literal * disables enforcement entirely. That is reasonable behind a reverse proxy, or with
a bearer token set, and not otherwise.
Rate limiting. A token bucket per client address, with a second bucket per MCP session
underneath it. PIHOLE_RATE_LIMIT (default 120, burst max(120/4, 30)) is the per-session
rate; one address may send four times that. Throttled requests return 429 with Retry-After: 1.
0 disables.
The address ceiling is what actually bounds a caller, because Mcp-Session-Id is a header the
client writes: without a ceiling, a caller sending a fresh session id on every request would never
be limited at all. The header is also being removed from the transport in MCP revision
2026-07-28, after which the ceiling is the whole limiter. Every request is charged to the
ceiling first, including one the session bucket then rejects, so a client that is already being
throttled keeps drawing on the budget its neighbours share. IPv6 clients are counted per /64,
since a single host is routinely given a whole one.
# Tighter limit for a small fleet
export PIHOLE_RATE_LIMIT=60
# Disable (only when running behind a proxy with its own rate limit)
export PIHOLE_RATE_LIMIT=0
Behind a reverse proxy. By default the client address is the connection's peer address, which
behind a proxy is the proxy itself, so every client through it shares one ceiling. Name the proxy
networks and X-Forwarded-For is used instead, taking the rightmost entry that is not itself a
listed proxy:
export PIHOLE_TRUSTED_PROXIES="192.168.1.5"
Nothing is trusted by default, deliberately: X-Forwarded-For is client-supplied, so a server
that believes it unconditionally lets any caller pick its own rate-limit bucket. Every host
inside a prefix you list gets that power, so name the individual proxies where you can and keep
any CIDR block as tight as the deployment allows. 10.0.0.0/8 trusts your whole network.
Tracing is opt-in. Set OTEL_EXPORTER_OTLP_ENDPOINT to enable:
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
pihole-mcp
All tool calls are automatically traced with tool name, duration, and error status.
If you don't need tracing, the slim build strips the OpenTelemetry SDK, gRPC, protobuf and grpc-gateway dependencies entirely. The binary is 43% smaller:
| Variant | Binary (linux/amd64) | Download (linux/amd64 .tar.gz) | Docker image (linux/arm64) |
|---|---|---|---|
| Default | 17.9 MB | 6.8 MB | 19.0 MB |
| Slim | 10.3 MB | 4.1 MB | 11.8 MB |
Measured from the v0.9.0 release assets. MB here is a million bytes, and the image figure is the unpacked size docker images reports for the linux/arm64 image, the one a Raspberry Pi pulls.
just build-slim
# or
go build -tags slim -o bin/pihole-mcp-slim ./cmd/pihole-mcp
# Docker
docker pull ghcr.io/hexamatic/pihole-mcp:latest-slim
The slim binary is functionally identical apart from OTEL_EXPORTER_OTLP_ENDPOINT being ignored.
Pi-hole allows a limited number of concurrent API sessions (webserver.api.max_sessions, 16 by default), and every client that logs in takes a seat: the web interface, PADD, Home Assistant, any other integration, and pihole-mcp. When they are all taken, Pi-hole answers 429 and refuses further logins, including from its own web interface.
pihole-mcp releases its seat on shutdown, but a session left behind by a process that was killed rather than stopped will hold one until it expires. Three ways out, in order of preference:
pihole_auth_sessions) and revoke one that is idle (pihole_auth_revoke_session).pihole-FTL --config webserver.api.max_sessions 32
webserver.session.timeout (30 minutes by default).Retrying will not help, so pihole-mcp does not: it reports the problem instead of silently stalling.
Pi-hole rate-limits repeated failed logins, and the limiter does not distinguish between "wrong password" and "the password you just fixed". Wait a few seconds and try again. If it persists, confirm you are using the admin password or an application password, not the web interface's TOTP code.
localhost inside a container is the container, not the host. Point PIHOLE_URL at the host's LAN address (http://192.168.1.2), at host.docker.internal on Docker Desktop, or put both containers on the same Docker network and use the Pi-hole container's name.
Every timestamp in tool output carries an explicit zone marker (e.g. 19 Jul 2026, 9:41 AM UTC), so responses are unambiguous whatever the zone. Which zone is used depends on where the server runs: native binaries use the system timezone, while the Docker image defaults to UTC. To get local times from the container, set TZ on the pihole-mcp container (not just the Pi-hole one). Timezone data is embedded in the binary, so no extra packages or volume mounts are needed:
environment:
- TZ=Australia/Adelaide
An unrecognised TZ value logs a warning at startup and falls back to UTC rather than refusing to start.
Your Pi-hole is serving HTTPS with a self-signed certificate, which fails standard TLS verification. The right fix is a trusted certificate on the Pi-hole (for example via its built-in domain settings or a reverse proxy with Let's Encrypt). If that isn't practical, set PIHOLE_TLS_SKIP_VERIFY=true to disable verification. Connections are still encrypted, but the server's identity is no longer checked, so only use this on a network you control.
Pi-hole's embedded web server closes connections under load. pihole-mcp retries these automatically with backoff; if you see failures anyway, raise PIHOLE_MAX_RETRIES (default 3).
See SUPPORT.md for the full picture, including where to start contributing.
# Prerequisites: Go 1.26+, Docker, mise, just
# One-command setup
just setup
# Start local Pi-hole (http://localhost:8081, password: test)
# Port already taken? PIHOLE_DEV_PORT=8091 just dev-up
just dev-up
# Run quality checks (format + lint + test)
just check
# Run integration tests against local Pi-hole
just integration
# Build binary
just build
See CONTRIBUTING.md for full development guidelines.
Pi-hole is a registered trademark of Pi-hole LLC. This project is independently maintained and is not affiliated with, endorsed by, or sponsored by Pi-hole LLC.
Source-derived launch command. Check the maintainer’s required arguments and credentials before running:
docker run -i --rm ghcr.io/hexamatic/pihole-mcp:0.9.0Merge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.
{
"mcpServers": {
"io-github-hexamatic-pihole-mcp": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"ghcr.io/hexamatic/pihole-mcp:0.9.0"
]
}
}
}Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.
Claude Desktop setup referenceghcr.io/hexamatic/pihole-mcp:0.9.0dockerPi-hole MCP Server works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.
~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.~/.cursor/mcp.jsonRestart Cursor for changes to take effect..vscode/mcp.jsonReload VS Code window for changes to take effect.~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect..mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.