Panelica MCP Server

Panelica MCP Server

Talk to your Panelica hosting panel in plain English. Provision a domain, issue an SSL certificate, create a database, or restart a service through Claude Desktop, Cursor, ChatGPT, or any other Model Context Protocol client.

404 tools cover the entire External API surface — accounts, domains, DNS, SSL, email, MySQL, FTP, security, backups, server resources, and more.

---

Table of Contents

• How it works
• Requirements
• Install
  • Option A — npm (recommended)
  • Option B — Docker
  • Option C — Build from source
• Configuration
  • 1. Open the External API port
  • 2. Generate an API key in the panel
  • 3. Verify the credentials with curl
• Wire it into your MCP client
  • Claude Code (one command)
  • Claude Desktop
  • Cursor
  • OpenAI Codex CLI
  • Continue.dev, Cline, Zed
  • Generic stdio client
• Tool catalogue
• Example sessions
• Permission scopes
• Security model
• Troubleshooting
• Development
• Versioning & support
• License

---

How it works

+----------------+         stdio JSON-RPC          +-------------------+
|  MCP client    |  <----------------------------> |  panelica-mcp     |
|  (Claude, ...) |                                 |  (this package)   |
+----------------+                                 +---------+---------+
                                                             |
                                          HTTPS + HMAC-SHA256|
                                            X-API-Key        |
                                            X-Timestamp      |
                                            X-Signature      |
                                                             v
                              +------------------------------+------------------------------+
                              |  https://<panel-host>:8443/api/external/v1/...               |
                              |  nginx reverse proxy on the panel host                       |
                              |  (TLS termination + path rewrite: /api/external/X -> /X)     |
                              +------------------------------+------------------------------+
                                                             |
                                       127.0.0.1:3002 plain  |
                                                             v
                                            +----------------+-----------------+
                                            |  external-server (HMAC verify)   |
                                            +----------------+-----------------+
                                                             |
                                                             v
                                            +----------------+-----------------+
                                            |  Panelica panel + Linux services |
                                            +----------------------------------+

panelica-mcp is a thin, stateless adapter:

1. The MCP client launches the binary over stdio.
2. At startup the server fetches <PANELICA_BASE_URL>/v1/api-spec from your panel and builds the tool list from it (live catalogue, default). If the panel is unreachable it falls back to the bundled tools/tools.json snapshot. Set PANELICA_LIVE_SPEC=0 to always use the snapshot.
3. The client asks for the tool list and receives it — together with server instructions (how ids, scopes, response envelopes, errors and rate limits work, plus workflow recipes), generated from that same catalogue. See What the assistant knows.
4. When the client calls a tool, the server builds the corresponding HTTP request, signs it with HMAC-SHA256 using your local PANELICA_API_SECRET, and forwards it to the panel.
5. The HTTP response is returned to the client as the tool result. A failed call comes back as an explanation of what to do next (missing scope, wrong id, rate-limit reset…) rather than a bare status code.

No data is cached, no telemetry is emitted, and the secret never leaves the machine running the MCP server.

Requirements

• A running Panelica panel (version 1.0.193 or newer recommended; the External API surface is stable from 1.0.180+).
• HTTPS access to the panel UI on port 8443 from the machine that will run panelica-mcp. This is the same port you already use in the browser — no extra firewall change is required.
• One of the following runtimes on that machine:
  • Node.js ≥ 20 for the npm install path
  • Docker for the container path

You do not need to install anything on the panel host itself, and you do not need to open the internal port 3002 to the public internet.

Install

Pick whichever fits your MCP client setup. All three produce the same stdio binary; pick by which sandbox model you prefer.

Option A — npm (recommended)

npm install -g panelica-mcp

or run without installing (the MCP client launches npx for you):

npx -y panelica-mcp

The -y flag accepts npm's "install on first run" prompt non-interactively, which is what MCP clients need.

Option B — Docker

A pre-built image is published to GitHub Container Registry on every release:

docker pull ghcr.io/panelica/panelica-mcp:latest

Run it from an MCP client config:

{
  "command": "docker",
  "args": [
    "run", "--rm", "-i",
    "-e", "PANELICA_BASE_URL",
    "-e", "PANELICA_API_KEY",
    "-e", "PANELICA_API_SECRET",
    "ghcr.io/panelica/panelica-mcp:latest"
  ],
  "env": {
    "PANELICA_BASE_URL": "https://your-panel-host:8443/api/external",
    "PANELICA_API_KEY":  "pk_...",
    "PANELICA_API_SECRET": "sk_..."
  }
}

-i keeps stdin attached so the MCP client can talk to the container. --rm removes the container when the client disconnects.

Option C — Build from source

git clone https://github.com/Panelica/panelica-mcp.git
cd panelica-mcp
npm install
npm run build
node dist/index.js          # speaks MCP over stdio

To regenerate tools/tools.json from your panel's live API spec:

PANELICA_SPEC_URL="https://your-panel:8443/api/external/v1/api-spec" npm run rebuild-tools

Configuration

You need three values: a reachable base URL, an API key, and an API secret.

1. Pick the right base URL

Panelica's external-server process listens on 127.0.0.1:3002, and the panel's nginx on 8443 reverse-proxies /api/external/... to it. Nginx strips the /api/external prefix before forwarding, so the path the HMAC signature is computed over and the path the backend sees both end up as /v1/... — signatures match end-to-end without any extra knobs.

The right PANELICA_BASE_URL depends on where you run panelica-mcp:

You should not open port 3002 to the public internet. The default install binds it on all interfaces but expects it to be either firewalled or only reached through the 8443 reverse proxy.

Sanity-check the proxy from your machine:

curl -sk https://<panel-host>:8443/api/external/health
# {"status":"ok"} or similar

If you get a TLS error, that is the panel's self-signed certificate — install a real cert on the panel (panel UI → Settings → SSL) rather than disabling verification client-side.

2. Generate an API key in the panel

1. Sign in to the panel as root or any account with permission to manage API keys.
2. Navigate to Settings → API Keys → Generate API Key.
3. Pick the scopes you want the MCP server to have. For a read-only assistant, *:read is enough (panel 1.0.528+; older panels need the individual <area>:read scopes). For full automation grant *:write and *:delete too, or *:*. Every tool's description in this server lists the scopes it requires.
4. Copy both key (looks like pk_...) and secret (looks like sk_...). The secret is shown only once; store it in a password manager.

3. Verify the credentials with curl

Before you wire the MCP client up, prove the credentials work end-to-end:

export PANELICA_BASE_URL=https://your-panel-host:8443/api/external
export PANELICA_API_KEY=pk_xxxxxxxx
export PANELICA_API_SECRET=sk_xxxxxxxx

TS=$(date +%s)
# Signature is over METHOD + PATH + TIMESTAMP + BODY. The path is the
# backend-visible path (/v1/...) — NOT the /api/external/ prefix that nginx
# strips before forwarding. panelica-mcp does this automatically.
SIG=$(printf "GET/v1/api-keys${TS}" \
  | openssl dgst -sha256 -hmac "$PANELICA_API_SECRET" -hex | awk '{print $2}')

curl -sk "$PANELICA_BASE_URL/v1/api-keys" \
  -H "X-API-Key:   $PANELICA_API_KEY" \
  -H "X-Timestamp: $TS" \
  -H "X-Signature: $SIG"

You should get back JSON listing your API keys. Common 401 responses:

4. Optional environment variables

Wire it into your MCP client

Claude Code (one command)

claude mcp add panelica \
  -e PANELICA_BASE_URL=https://your-panel:8443/api/external \
  -e PANELICA_API_KEY=pk_... \
  -e PANELICA_API_SECRET=sk_... \
  -- npx -y panelica-mcp

That's it — ask Claude Code to "list my domains" or "create a database for example.com" and it will drive the panel through the scoped API key.

Claude Desktop

Edit your Claude Desktop config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux (Claude Desktop beta): ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "panelica": {
      "command": "npx",
      "args": ["-y", "panelica-mcp"],
      "env": {
        "PANELICA_BASE_URL":   "https://your-panel-host:8443/api/external",
        "PANELICA_API_KEY":    "pk_...",
        "PANELICA_API_SECRET": "sk_..."
      }
    }
  }
}

Save, fully quit Claude Desktop (not just close the window — Quit), and re-open it. A new chat will show panelica as a connected MCP server with "37 tools available" (the core set plus three meta tools; set PANELICA_TOOLSETS=all for all 422).

Cursor

In Settings → MCP → Add new server:

{
  "panelica": {
    "command": "npx",
    "args": ["-y", "panelica-mcp"],
    "env": {
      "PANELICA_BASE_URL":   "https://your-panel-host:8443/api/external",
      "PANELICA_API_KEY":    "pk_...",
      "PANELICA_API_SECRET": "sk_..."
    }
  }
}

OpenAI Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.panelica]
command = "npx"
args = ["-y", "panelica-mcp"]

[mcp_servers.panelica.env]
PANELICA_BASE_URL = "https://your-panel:8443/api/external"
PANELICA_API_KEY = "pk_..."
PANELICA_API_SECRET = "sk_..."

Continue.dev, Cline, Zed

Any MCP-aware editor that accepts a stdio command works the same way — give it npx -y panelica-mcp (or the absolute path to the built dist/index.js) and the three environment variables.

Generic stdio client

PANELICA_BASE_URL=https://your-panel-host:8443/api/external \
PANELICA_API_KEY=pk_... \
PANELICA_API_SECRET=sk_... \
panelica-mcp

The process speaks MCP JSON-RPC over stdin/stdout. Send an initialize request first, then tools/list, then tools/call.

Toolsets — what the client sees

MCP clients budget tools: Cursor caps active tools at about 40 across all servers and silently drops the rest, and every registered tool costs prompt tokens on each turn. So by default the server registers a compact core set (34 everyday tools: accounts, domains, DNS, SSL, databases, e-mail, FTP, backups, server status/services, WordPress, Docker, plans) plus three meta tools that reach the whole catalogue:

Choose the set with PANELICA_TOOLSETS (comma-separated, unioned):

Direct calls to a non-registered catalogue tool are still accepted, so a client that learned a tool name from panelica_find_tools can call it either way.

What the assistant knows

An assistant that only sees 400 tool names guesses ids, retries 403s and misreads responses. panelica-mcp gives it the API "by heart" instead, and everything below is generated from the catalogue, so it can never promise a path the panel does not have:

• Server instructions (sent on initialize, so every client injects them into the model's context): how calls work, that every id is a UUID coming from a list call, that users are called accounts, the {"status":"success","data":…} envelope, unpaginated lists, what each error class means and what to do, scope families, the category map with counts, safety rules for mutating/destructive tools, and workflow recipes that are included only when every step exists in the catalogue.
• Your key, up front. At startup the server calls GET /v1/me once and tells the model which key it holds, its scopes, tier and expiry — so a read-only key is announced as read-only and out-of-scope tools are declined with the scope to add, not attempted. A rejected credential is reported in the instructions before the first tool call. (PANELICA_STARTUP_PROBE=0 disables the probe.)
• Scopes as the code enforces them. Newer panels derive the required scopes from the middleware chain instead of hand-written docs, including "one of a, b" rules and role restrictions (ROOT/ADMIN); the tool description repeats them verbatim.
• Allowed values. Fields with a declared value list (ssl_provider, web_server, restart_policy …) carry a JSON-schema enum and default, so the model does not guess.
• Big lists stay usable. Every list-returning GET accepts _limit, _fields and _match; the server applies them to the response (the panel never sees them) and reports {total, matched, shown} so a 500-domain panel can be searched in one call.
• Validation errors name the field. A Gin validation failure becomes field "user_id" is required / field "type" must be one of the allowed values using the JSON names the model used.
• Search speaks the user's language. panelica_find_tools maps website→domain, certificate→ssl, mailbox→email, container→docker and stems plurals, so the first query finds the right tool.
• Transient limits are absorbed. A 429 whose window resets within 15 s is waited out once; a GET that fails on the network is retried once; a clock skew above two minutes against the panel is flagged in the instructions (HMAC timestamps would be rejected).
• Per-tool descriptions carry the HTTP route, required scopes, a Returns: line with the response envelope and the first data fields, and the risk class (read-only / mutating / destructive).
• Id parameters say where they come from — UUID of the domain — obtain it from GET /v1/domains — for path, query and body fields alike. Newer panels ship this in the spec; for older panels the server derives it.
• panelica_describe_tool returns the exact parameters, body fields and response fields of one tool, so the model checks before it calls.
• Errors are explained, not echoed: a 403 names the missing scope and says retrying cannot help; a 404 says to re-list and use a real id; a 429 reports the reset window; 5xx says to report, not loop. The raw panel response is appended for reference.
• Results are bounded: oversized list responses are cut to the first items with an explicit _truncated note (still valid JSON) instead of flooding the context.
• Resources for clients that support them: panelica://guide (the instructions), panelica://catalogue (every tool with route and summary) and, in live mode, panelica://spec (your panel's full API spec).

The response-field lists come from the panel's /v1/api-spec, which newer panels derive from their handler source at build time; a panel that predates this still works — its tools simply show {status, data} without field names.

Tool catalogue

422 tools are auto-generated from the panel's live /v1/api-spec, so they never drift from the API. Each tool carries MCP safety annotations ( 181 · 177 · 46) that capable clients use to auto-approve reads and warn before destructive calls. Body fields and query parameters are derived from the panel's handler code (gen-external-schemas in the panel repository), so on panels from 1.0.528 on mutating tools carry typed, required-marked bodies and action endpoints such as …/suspend or …/restart declare that they take no body at all.

Full machine-readable list: tools/tools.json.

Example sessions

After wiring the server up, try these in your MCP client:

Domain provisioning.

"Create a new account for alice@example.com on the starter plan, then add the domain alice-shop.com to it and issue a Let's Encrypt certificate."

The assistant will pick up panelica_accounts_post_v1_accounts, panelica_domains_post_v1_domains, and panelica_ssl_post_v1_ssl_... from the catalogue, fill in the parameters from the conversation, and call them in sequence. You can watch the calls happen in the client's tool log.

Diagnostic.

"Show me the last 24 hours of bandwidth usage for alice-shop.com and tell me whether it is on track to exceed the plan quota this month."

Bulk cleanup.

"List every domain whose SSL certificate expires in the next 14 days and renew them all."

DNS migration.

"For the zone alice-shop.com, list the current A and CNAME records, then add www as a CNAME to alice-shop.com and an A record for mail pointing to 203.0.113.10."

The assistant will only invoke tools whose scopes are granted to your API key, so a read-only key safely answers "list" questions but refuses "create / delete".

Permission scopes

API keys are scoped — grant an AI assistant exactly the access it needs, nothing more. No scope is preselected when creating a key in the panel, and the create dialog has live search over all 50 scopes. view/list only · create/update · remove · special access. Every family also accepts its wildcard (domains:*) and *:* grants everything.

Mutating service control deliberately requires its own action scopes (or server:write) — a metrics-only server:read key can not stop MySQL.

Rate limits

API keys carry a rate-limit tier (rate_limit_tier when creating the key: starter 60/min, professional 300/min, business 1000/min, enterprise unlimited). The default starter tier is easy to exhaust when an assistant fans out many calls; create the assistant's key with a higher tier if you see 429 RATE_LIMIT_EXCEEDED (the response carries reset_at).

Security model

• HMAC-SHA256 request signing. Every request is signed over METHOD + PATH + QUERY + TIMESTAMP + BODY with your API secret. The panel rejects requests whose timestamp drifts more than 5 minutes from server clock, so replays are not possible.
• Secrets stay local. The API secret is read from the process environment and used only to compute the signature. It is never logged, sent to any third party, or written to disk.
• Scope-restricted keys. Generate one API key per use case. Grant only the scopes that use case needs — e.g. domains:read for a read-only assistant, *:write only for full automation.
• Audit trail. Every request hits the panel's normal audit logging and RBAC. Actions taken via MCP are indistinguishable from any other authenticated API call and can be traced to the API key that performed them.
• No data harvesting. This server emits no telemetry, writes no cache, and contacts no third party.
• Container hardening. The Docker image runs as a non-root user and exposes no ports — it speaks only stdio.

Troubleshooting

If you are still stuck, open an issue at github.com/Panelica/panelica-mcp/issues with the (redacted) stderr output.

Development

git clone https://github.com/Panelica/panelica-mcp.git
cd panelica-mcp
npm install
npm run build
node dist/index.js

Project layout:

.
├── src/index.ts          # MCP server (stdio transport, HMAC client, live-spec loader)
├── src/catalog.ts        # spec → tools generator (shared by runtime and build script)
├── src/toolsets.ts       # core set, PANELICA_TOOLSETS selection, find/describe/call meta tools
├── src/instructions.ts   # server instructions generated from the catalogue
├── src/errors.ts         # error explanations + result bounding
├── tools/
│   ├── build-tools.mjs   # Generates tools.json from the API spec
│   ├── api-spec.json     # Committed snapshot of the live /v1/api-spec
│   └── tools.json        # tool definitions, auto-generated (committed)
├── .github/workflows/
│   └── refresh-tools.yml # Weekly CI: regenerate, bump, tag, publish npm + Docker + Registry
├── Dockerfile
├── smithery.yaml         # Smithery deployment manifest
├── .env.example
└── README.md

Keeping the tool catalogue current

The catalogue never drifts from the API by hand. The backend serves an always-current /v1/api-spec (built from its route registry), and the tools are regenerated from it:

# Rebuild from the committed snapshot (offline):
npm run rebuild-tools

# Pull the live spec, refresh the snapshot, and rebuild:
PANELICA_SPEC_URL="https://your-panel:8443/api/external/v1/api-spec" npm run rebuild-tools

CI (refresh-tools.yml) runs this weekly against the panel named in the PANELICA_SPEC_URL repository secret. When the catalogue changed it bumps the patch version, tags, and publishes npm, the Docker image and the MCP Registry entry in the same run, so a new API endpoint reaches npx -y panelica-mcp users automatically. Running servers do not even need that: with the default live catalogue they read the spec from the panel they talk to. Each tool is tagged with safety annotations (readOnlyHint for GET, destructiveHint for DELETE) that capable MCP clients use to auto-approve reads and warn before destructive calls.

A separate, internal dataset of every panel endpoint (1,263 total) exists for training purposes — only the 404 documented External API endpoints are exposed through this package. Internal panel endpoints, recorded DEV data, and training jsonl files are not part of the public repository.

Versioning & support

• This package follows the Panelica panel's External API. Tool signatures change only when the panel itself ships a backward-incompatible API change, and the package's major version is bumped to match.
• New endpoints become available the next time we regenerate tools/tools.json and publish a release — or immediately, in live mode, from your own panel's /v1/api-spec.
• Response-field lists, id provenance and cleaned descriptions come from the panel's spec; a panel that predates them still works with plainer tools.
• Panel issues (the API itself, not this client): the Panelica forum at forum.panelica.com.
• Client / packaging issues: github.com/Panelica/panelica-mcp/issues.

License

MIT. See LICENSE.

Links

• Panel website: https://panelica.com
• Live demo: https://demo.panelica.com
• Documentation: https://panelica.com/docs
• Forum & support: https://forum.panelica.com
• Public installer: https://latest.panelica.com/install.sh