Lets coding agents run dev servers and workers as real systemd, launchd or Windows services, and read their logs.
Agents love backgrounding things with & or nohup. Then the session ends, the dev server dies or turns into an orphan, and its output is gone. The SPM MCP server offers a sturdier habit: start that process as a genuine native service, let the operating system restart it, and read its logs later through structured tool calls.
SPM wraps whatever the platform already has rather than inventing its own daemon model.
| Platform | Backend | Default user-scope storage |
|---|---|---|
| Linux | systemd and journald | ~/.config/systemd/user |
| macOS | launchd LaunchAgents, or LaunchDaemons with --system |
Application Support |
| Windows | Per-user supervisor and Startup entry, or SCM services with --system |
%APPDATA%\spm |
User services need no elevation. Machine-wide services require root or an Administrator terminal, and on Windows can run under LocalService, NetworkService, a domain account or a managed service account. A single spm start ./my-server --name my-server creates the service, starts it, enables it at login and restarts it on crash. Scripts in Node, Python, Ruby, shell and PowerShell are recognised by extension.
The MCP side splits into read-only tools for listing, inspecting, diagnostics and bounded log reads, and a smaller set of write tools for creating and controlling services. Anything else in the CLI is reachable through spm_run_cli: schedules, signals, scaling, declarative apply and diff, import and export.
Two design choices stand out. Unbounded calls such as following logs or watching the list are refused, since an MCP call has to return. And both spm_run_cli and service deletion are flagged destructive, so a client can ask you before running them. Secrets belong in an env file referenced by the service, not in tool arguments.
Install the binary first with the Linux and macOS shell installer or the Windows PowerShell installer; both verify release checksums. Building from source needs Go 1.27 or later. Then start the SPM MCP server with spm mcp. Claude Code users can add it directly or install the plugin, which bundles a skill that teaches agents to use spm and to ask before deleting services. Claude Desktop may start with a minimal PATH, so an absolute path to the binary can be necessary.
Linux requires systemd. Health timers, instances, resource limits and unit import are Linux-only for now. On Windows and macOS, log reads support line counts and grep but not time windows. PM2-style cluster mode, watch mode and remote management are not implemented, and runtimes like Node are not installed for you. See more process and system bridges in other integrations, our integrations index for tools outside the main categories.
Check the linked publisher documentation and your connected client's tool list. Names and availability can vary with the server version and account permissions.
| Tool | What it does |
|---|---|
| spm_run_cli | Runs other non-interactive spm CLI commands, including schedules, signals, scaling, declarative apply and diff, import and export. Marked destructive so clients can request approval. |
{
"mcpServers": {
"spm": {
"command": "spm",
"args": [
"mcp"
]
}
}
}Install spm first. If Claude Desktop cannot find spm, use the absolute path to spm or spm.exe, then fully restart the app.
Run an Actual Budget ledger from your AI agent, with phone approvals for every change on paired machines.
An Odoo module that serves a read-only MCP endpoint at /mcp, scoped to each user's own access rights.
Agent memory that keeps every correction with its history, plus saved working state so a crashed run can resume.
Targeted .NET assembly inspection for AI agents on Windows: find types, decompile one method, trace callers.
Manage a Linux VPS, Postgres and Odoo databases and GitLab deploys over SSH, with previews, allowlists and an audit trail.
Ask an assistant for today's internet-wide cyber threat band, score, critical advisories and data-source health.