Potencia tu agente
La URL propia del publicador. No es un endpoint de PlugMCP.
Host a copy with our keys and logs.
https://{drop_host}/api/v1/mcpEl botón de Claude de arriba abre Claude con el nombre y la URL ya rellenados. Aún tienes que confirmarlo allí.
claude mcp add --transport http io.github.julesnsenda-drop https://{drop_host}/api/v1/mcpRun in a terminal.
{ "mcpServers": { "io.github.julesnsenda-drop": { "command": "npx", "args": [ "-y", "mcp-remote", "https://{drop_host}/api/v1/mcp" ] } }}claude_desktop_config.json
¿URL remota o un comando en tu máquina? En qué se diferencian las dos instalaciones
Ejecutar en PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
AbrirDel publicador
Deploy, Run, Operate, Publish
A lightweight, self-hosted Platform as a Service (PaaS) engineered for the "drop folder and deploy" workflow. Zero-configuration deployment for Node.js, Python, Go, static sites, and containerized applications.
Drop a folder, get a URL. Zero configuration for 80% of use cases.
Self-hosted PaaS: deploy apps from files or git, read logs, check status, roll back.
DATABASE_URL injecteddrop.yaml/api routingDROP_DATA_DIRFull documentation lives at dropkit.sh/docs, and every CLI command and API endpoint is catalogued in the reference. This README covers installing DROP, the trust model you should read before exposing it, and disaster recovery.
Also in this repo: HTTPS setup, git redeploy and custom domains, agent deploys.
Release install (recommended, Linux): a fresh Debian/Ubuntu box with root access. install.sh provisions Node.js, PostgreSQL, Caddy, and (if you choose --isolation=docker) Docker Engine for you — no toolchain to install yourself.
Build from source: Node.js 20+, npm 9+, and optionally Caddy 2.0+ for hostname routing.
install.sh refuses to run piped straight from curl — it needs an on-disk copy of itself to work from — so save it first, then run it:
curl -fsSL https://github.com/JulesNsenda/drop/releases/latest/download/install.sh -o install.sh && sudo bash install.sh --from-release --isolation=docker
--from-release installs the latest published release: no git clone, and no TypeScript or Vite build on your machine. (Node.js is still required and the installer sets it up for you. npm may log a failed optional build for cpu-features — that is harmless; it is an optional native accelerator for ssh2 and the install completes without it.)
It requires you to pick an isolation mode explicitly on a first install:
--isolation=docker (recommended) — tenant apps build and run in Docker containers with resource limits and no access to the platform's secrets.--isolation=none — tenant apps run as the drop system user, the same user that owns the platform's encryption key and JWT/OAuth secrets. Only choose this on a single machine where every deployer is fully trusted; see Security & Trust Model.Add --domain=example.com --https [email protected] once DNS points at the box, or pin a specific version with --from-release=v1.0.0 instead of the latest.
Once the service is up, retrieve the one-time admin password from the platform log and change it immediately:
journalctl -u drop-platform -b --no-pager | grep -A1 'Default Admin Credentials'
Windows:
install.shtargets systemd/apt Linux boxes. On Windows, runinstall.bat(or build from source) and start withdrop servedirectly — Windows is fully supported underisolation: none.
install.sh verifies the downloaded tarball's SHA-256 checksum itself before installing anything. To also verify it was built by this repo's own GitHub Actions, check the build provenance attestation:
gh release download --repo JulesNsenda/drop -p drop-dist.tar.gz
gh attestation verify drop-dist.tar.gz --repo JulesNsenda/drop
Attestations are attached by the release workflow on every release built after this repository became public. If
gh attestation verifyreports no attestation for a given release, that release predates it — fall back to the publisheddrop-dist.tar.gz.sha256, whichinstall.shchecks automatically.
Every release also attaches drop-dist.tar.gz (compiled server, CLI and dashboard), its .sha256, and install.sh, so you can fetch and inspect them by hand on an air-gapped box. All releases are listed here.
For contributors, or to run from a branch instead of a release:
git clone https://github.com/JulesNsenda/drop.git
cd drop
npm install
(cd src/dashboard && npm install) # the dashboard is a separate package
npm run build # compiles the server AND builds the dashboard
npm link # makes the 'drop' command available globally
If you only changed backend code,
npm run build:serverskips the dashboard build.
On first start, drop serve prints the one-time admin password to the console. Change it immediately.
drop serve # start the platform
cp -r my-app /var/drop/data/webapps/ # drop a folder in (Windows: xcopy to C:\drop\data\webapps\)
DROP detects the app type, installs dependencies, builds, provisions a PostgreSQL database if the app needs one, and starts it on an assigned port. Your app is then at http://localhost:<port>, or at http://my-app.localhost with Caddy installed. Edit any file and it rebuilds and restarts.
The dashboard is at http://localhost:3000/dashboard — apps list, per-app start/stop/restart, secrets, custom domains, deploy history, logs, user management, and the Claude connector details.
For the CLI and REST API, see the reference. A running DROP redirects /dashboard/docs and /dashboard/reference there.
Read this before exposing DROP to anyone you don't fully trust.
DROP ships two explicit isolation modes with different trust guarantees.
isolation: none (default) — single-user / trusted deploymentsDeploying an app means running its code (install/build scripts and the app process) on the host as the DROP process user. A deployed app — or its build script — can read other apps' data and the platform's own files.
Use this when: it's your machine or a machine you control, and everyone with deploy access is trusted. Treat deploy access like shell access.
allowSignup in this mode — DROP refuses at startup.DROP_DISABLE_AUTH=true).isolation: docker — multi-user / invited usersApps build and run in Docker containers with strict resource limits (--cap-drop=ALL, --security-opt no-new-privileges, memory/CPU caps, --pids-limit). Build containers have no access to platform secrets and cannot reach the LAN or cloud-metadata endpoints.
Honest residual risks (documented, not hidden):
a.yourdomain.com and b.yourdomain.com can read each other's cookies. Use a dedicated baseDomain for multi-tenant use, or submit it to the Public Suffix List.allowSignup: true) enables self-service registration. Abuse tooling, takedown runbooks, and egress enforcement for hostile public access are future work. Treat open-internet signup as documented residual risk until then./app bind mount (no image commit), so only dependencies written into the app dir reach the runtime. Node (node_modules), Go (compiled binary) and Python (an /app-local .venv, whose bin/ is put on the runtime PATH) all land there. Anything a custom build command installs into system site-packages or a global prefix is discarded with the build container — the build "succeeds" and the app then fails to import at boot.Requires Docker Engine on Linux (Docker Desktop on Windows/macOS is dev/best-effort only for this mode).
Build containers run as the platform's own (non-root) user so they can write the app source without needing CAP_DAC_OVERRIDE. Apps deployed through DROP (git deploy, webhook, drop deploy) are owned by that user automatically. If you instead place a folder into data/webapps/ by hand, own it as the platform user (e.g. chown -R drop:drop) — a folder owned by a different user (a sudo cp as root) will fail the build with EACCES (fail-closed by design; DROP will not run an untrusted build as root to work around it).
readonly/user/admin)x-forwarded-for)drop.yaml schema (unknown keys rejected; TLS paths confined to app dir)drop.yaml values never reach docker run args or mount specs directly.git metadata are refused outright.git/config or visible in psSee .env.example for all security-relevant settings.
DROP keeps critical state in the file stores under data/drop-svc/ (credentials, encrypted secrets, the encryption key, webhooks, app state) and in PostgreSQL — both the internal drop_internal database and every provisioned per-app database. Snapshot all of it with:
drop backup # writes data/backup/backup-<timestamp>/
drop backup --keep 14 # keep the newest 14, prune the rest
A backup contains the JSON/YAML stores, encryption.key, a pg_dump of drop_internal, and a pg_dump -Fc of each per-app database under databases/ (plus a generated databases/restore-roles.sql that recreates the app DB roles). Schedule it yourself (cron / Task Scheduler) — DROP does not run backups automatically. The backup command exits non-zero if any dump — per-app, internal, or the database enumeration itself — fails, so a cron job that only checks the exit code will still catch a partial backup.
Caveat: backups are same-platform only. A backup taken on Windows will not restore on Linux (and vice versa) — the bundled PostgreSQL binaries and data layout are platform-specific.
drop restore reverses a backup. It is destructive — it overwrites the current file stores and databases — so it refuses to run while the platform is up, requires --confirm, and prints its plan first:
drop restore data/backup/backup-<timestamp>/ # prints the plan, writes nothing
drop restore data/backup/backup-<timestamp>/ --confirm # actually restores
Stop the platform first. A running drop serve holds state in memory and would stomp the restore; drop restore refuses if it detects a daemon or a foreground platform still answering on the API port.
File stores (data/drop-svc/, data/appconf/webapps/) are copied back with their modes preserved (secrets stay 0600).
Databases are replayed with the bundled psql/pg_restore — but note that drop server stop also stops the bundled PostgreSQL, so in the normal flow drop restore finds it unreachable, prints the exact per-database commands, and skips the automatic DB step. To restore databases automatically, start PostgreSQL standalone first and re-run:
"<dropRoot>/apps/drop-svc/pgsql/bin/pg_ctl" -D "<dropRoot>/data/db" start
drop restore data/backup/backup-<timestamp>/ --confirm
The DB step authenticates against the currently running server's data/drop-svc/.pg-superuser (read before any file is overwritten), not the backup's copy. After a restore the running server's password and the restored file may diverge until the next platform restart.
Doing it by hand (equivalent to what drop restore runs, and what it prints when it skips the DB step):
BIN=<dropRoot>/apps/drop-svc/pgsql/bin ; export PGPASSWORD="$(cat data/drop-svc/.pg-superuser)"
# 1. Recreate app roles (clean server runs clean; on a re-run, "role already exists" is expected/benign)
"$BIN/psql" -h 127.0.0.1 -p 5433 -U postgres -d postgres -f databases/restore-roles.sql
# 2. Recreate + restore each database, drop_internal included (--create makes the DB AND restores REVOKE CONNECT)
"$BIN/pg_restore" -h 127.0.0.1 -p 5433 -U postgres --create -d postgres databases/drop_<app>.dump
# (re-run over existing DBs: add --clean --if-exists ; check exit codes, don't ignore stderr)
Use the bundled pg_restore/psql under apps/drop-svc/pgsql/bin — not a system Postgres client, since major-version mismatches can corrupt the restore. Backups are same-platform only (a Windows backup won't restore on Linux), and the DB restore round-trip is not covered by automated tests — validate on a non-production box first.
Deleting an app (drop remove <app> / DELETE /api/v1/apps/:name) dump-then-drops its provisioned database: before the database is dropped, DROP pg_dumps it to data/backup/pre-delete/<db>-<timestamp>.dump plus a companion <db>-<timestamp>.restore-role.sql (recreates the role, since -Fc doesn't capture it). The drop only happens if the dump verifies; if pg_dump fails or Postgres is down, the database is left intact instead of being lost. Pre-delete dumps are retained for DROP_PREDELETE_RETENTION_DAYS days (default 3) and pruned automatically on each subsequent delete — copy any you want to keep permanently off-box before then.
Run drop remove --keep-data <app> to skip dump-then-drop entirely and leave the database in place.
To restore a pre-delete dump, use the same procedure as the Restore section above: run its .restore-role.sql with psql, then pg_restore --create the .dump file.
DROP keeps PM2-managed app processes running across a platform restart, but the bundled PostgreSQL and Caddy are stopped and restarted with the platform, so expect a brief blip in database connectivity and routing during an upgrade.
drop backup).drop server stop then drop serve -d after upgrading — a plain pm2 restart keeps the old args/path from the PM2 dump.drop serve -d honors --root/--domain/--https/... flags that were previously ignored. If you have been passing flags that had no effect, they now take effect — review them before upgrading (e.g. a stray --https will actually enable HTTPS and validate your domain config).Breaking changes are listed at the top of each release's notes in the CHANGELOG.
npm run dev # start in development mode
npm run build # build server + dashboard
npm test # run tests
npm run lint # lint
Branching model and conventions: docs/GIT-BRANCHING-MODEL.md. Roadmap detail: docs/VERSION-ROADMAP.md.
MIT License - see LICENSE for details.
Self-hosted PaaS: deploy apps from files or git, read logs, check status, roll back.
{ "mcpServers": { "io.github.julesnsenda-drop": { "url": "https://{drop_host}/api/v1/mcp", "headers": {} } }}Herramientas
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Copia en esta página la configuración http para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.
Este registro no indica ninguna variable de entorno obligatoria.
Se accede a DROP por HTTP.
Última comprobación: 2026-11-09. Self-hosted PaaS: deploy apps from files or git, read logs, check status, roll back.