Some hosts can't enforce cgroup limits for nested containers at all
(confirmed on an OpenRC-based unprivileged LXC with no systemd to
delegate cgroup controllers) — docker run fails outright trying to
write memory.max. 0 now skips requesting that cap instead of erroring
every deploy.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The manager image had git but not openssh-client, so ssh:// clones
failed outright ("cannot run ssh"). Install openssh-client, and mount
an SSH_DIR (key + known_hosts) into the container read-only so private
repos can actually authenticate — confirmed both gaps while deploying
against real private Zenvairo repos.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
CONTAINER_MEMORY_MB and CONTAINER_CPUS (defaults 512/1) are now applied
to every installed MCP container's HostConfig instead of leaving them
unbounded.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Replaces the slice-2 hardcoded installation with a full install/deploy
pipeline:
- SQLite persistence (better-sqlite3) with a single `installations`
table and a draft -> cloning -> building -> built -> starting ->
running/error/stopped state machine; env vars and bearer tokens are
encrypted at rest (AES-256-GCM, key derived from MASTER_KEY).
- Build/run orchestration: git clone (+ .env.example key detection),
`nixpacks build` as a logged subprocess, dockerode container
run/stop with forced host/port env vars and a TCP-connect health
check before flipping an installation to "running".
- REST API (installations CRUD, build/deploy/redeploy, logs, token
regeneration) behind single-operator session-cookie auth with a
basic login lockout.
- Vite+React admin SPA: login, installation list, install wizard, and
a detail page to configure env vars, trigger build/deploy/redeploy,
view the build log, and copy the authenticated /mcp/<slug> URL.
- Dockerfile (multi-stage, pulls in nixpacks + the Docker CLI) and
docker-compose.yml for production deployment.
Validated end-to-end against a real cloned repo: draft creation with
env-key detection, build, the error path on a failed health check, and
delete, all through the live API.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Lays the groundwork for remote-mcp-manager: an Express server that
runs installed MCP containers via dockerode (forcing host/port env
vars, no published ports, shared bridge network) and reverse-proxies
/mcp/:slug requests behind a per-installation Bearer token, resolving
the upstream target by inspecting the container's network IP rather
than relying on container-name DNS.
Validated end-to-end against a real cloned MCP repo (nixpacks build +
container run + authenticated proxy roundtrip); installations are
still a hardcoded in-memory list, replaced by SQLite in the next slice.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
