Services

crow's infrastructure runs as daemons supervised by crow-cli daemon:

service	what it is	health
crow-memory	Rust HTTP service over LanceDB — sessions, messages, multivector ColBERT search	/healthz
crow-mcp	the memory tools (and more) as MCP over HTTP	tcp port
ollama-mv	multivector embedding server (the colbert: true model)	/api/version
searxng	docker container backing web_search	/
anything in services:	e.g. playwright MCP	tcp port from its mcpServers url

Conventions: pidfile at ~/.agents/crow/run/<name>.pid, log at ~/.agents/crow/logs/<name>.log (rotates 5 MB × 4).

The commands

crow-cli daemon list          # pid / running / healthy / unmanaged for all
crow-cli daemon start all     # no-op if already running
crow-cli daemon stop crow-mcp # never kills processes we didn't start
crow-cli daemon restart all
crow-cli daemon install ollama-mv   # build what's missing, point config at
                                     # it, start, verify. Idempotent.

Unmanaged is sacred. If a service is up but we don't hold its pidfile (you started it by hand, systemd resurrected it), start/stop/restart refuse to touch it. That's deliberate — check status first.

services: — run anything as a daemon

The top-level services: block in config.yaml declares how to run things — typically MCP servers — as daemons. Each entry takes the daemon fields (command, args, env, health_url, tcp_port, start_timeout, stop_timeout, or the docker ones: kind, container, compose_file, compose_service):

mcpServers:
  playwright:
    transport: http
    url: http://localhost:2779/mcp

services:
  playwright:
    command: npx
    args: ['@playwright/mcp@v0.0.79', --port, '2779']

A service that shares its name with an mcpServers: entry and declares no health check gets its tcp probe port from that entry's url — one source of truth for the port. Then crow-cli daemon start all brings up the whole stack, and agent sessions connect over HTTP instead of spawning stdio servers per session.

Precedence: built-ins < services: < daemons: (per-field overrides, and daemons: always wins).

Why not systemd?

It used to be systemd user units; the supervision now lives entirely in crow-cli daemon so it works identically on any machine with the repo. If something respawns "magically" after a kill, check systemctl --user list-units | grep ollama for a ghost unit from an old setup.