
Drobek
io.github.freemav0.6.0Updated Sep 29, 2026
Build, preview, version and publish small browser apps from coding agents over MCP.
Overview
Lets a coding agent build, preview, version and publish small browser apps on a Drobek hosting instance.
- What it does
- Drobek is an open-source hosting platform for small web apps built by an AI agent. Over MCP the agent calls tools such as create_app, write_files, read_file, publish and restore_version: it writes app sources into a workspace, the server compiles them with esbuild and returns compile errors in the same response, and each successful compile appears on a preview host. Publishing moves a pointer to an immutable version. Platform modules add sign-in, data collections, forms, e-mail, uploads and an SSRF-guarded proxy for external APIs.
- When to use it
- Use it when you want an assistant to turn a small idea into a working, shareable browser app without you setting up a project, build pipeline or hosting. It suits prototypes, internal tools and small public apps that need sign-in, stored records, forms or uploads. It is not meant for arbitrary server-side code or per-app npm installs.
- Requirements
- The hosted instance is a remote Streamable HTTP endpoint at drobek.app/mcp using OAuth 2.1, with scopes read, write and publish; no browser on the agent machine means minting an API key and sending it as an Authorization Bearer header. Self-hosting needs a linux/amd64 server with a public IP and ports 80 and 443, a dashboard domain plus an apps domain with wildcard DNS, Docker, and an SMTP account or Resend API key for sign-in codes.
Installation
In SourceWeft
- Open Drobek in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.
Other MCP clients
Add this to your client's mcpServers config.
{
"mcpServers": {
"drobek": {
"type": "http",
"url": "https://drobek.app/mcp"
}
}
}README
drobek
[CI] [Release] [License: AGPL-3.0] [Image]
Open-source hosting for small web apps your own AI agent builds.
Bring Claude, Claude Code, Cursor or Codex. Your agent connects over MCP, writes the app directly into a workspace and gets compile feedback and a live preview. You decide when to publish. Platform modules supply sign-in, data, forms, e-mail, uploads and external APIs; a dashboard gives you control over the app, its users and its secrets.
Try drobek.app · Self-host · Gallery · Docs
MCP Registry name: io.github.freema/drobek. The hosted
Streamable HTTP endpoint is https://drobek.app/mcp and uses OAuth 2.1.
Choose where to run it
The core is AGPL-3.0. Building, testing and self-hosting need no access to the private repository that operates drobek.app. The agent plugins are MIT and can connect to either instance.
See what people build
- Drobek Tycoon — a playable hosting-company simulator with three levels.
- Pokédex — PokéAPI search and stats through drobek's proxy module, with a live request log.
- Pixel Wall — a shared 64 × 40 canvas backed by auth, data and other platform modules.
- Pixel Crumbs — a browser pixel-art and animation editor.
- Drobek Skok — a pixel platformer with a leaderboard.
- Pekárna Drobek — a small bakery's demo website.
These are public apps on drobek.app. Browse the gallery for more examples.
[Drobek Tycoon running on drobek]
Build your first app
-
Choose an instance above and connect your agent.
-
Ask for a small app, for example:
-
Open the preview, ask for changes, then ask the agent to publish when ready.
Apps run in the browser. Their backend comes from installed platform
modules; arbitrary app server code and per-app npm install are outside
the execution model. For your own backend integration,
write a platform module.
The loop
- Connect drobek to your agent (an MCP server with OAuth — you approve it in the browser).
- Ask for an app: "a shift planner for our warehouse". The agent calls
create_appand gets a compiling starter plus a briefing of the rules. - It calls
write_files. drobek compiles the files in-process with esbuild and returns the compile errors in the same response; the agent fixes them and writes again. Every write is an immutable version. - Each successful compile is live at once on the app's preview host,
https://<slug>--preview.<APPS_DOMAIN>— the agent hands you the link. - When you are happy, the agent calls
publishand the version goes live onhttps://<slug>.<APPS_DOMAIN>(or your own domain). Publishing an older version is the rollback.
Why it is built this way:
- The server never executes app code. It compiles and serves; the result
runs only in browsers. No sandbox per app, no
npm install, no server-side code of the agent's. - Secrets never pass through the agent. The app owner sets them in the dashboard; modules use them server-side.
- Every app is its own origin (
<slug>.<APPS_DOMAIN>), separate from the dashboard's. - One process, one image (
ghcr.io/freema/drobek) + Postgres + Redis (+ Caddy for TLS). It runs on an ordinary small server.
Core concepts
Agent plugins and platform modules do different jobs. A plugin teaches your agent how to build on drobek. A module runs on the drobek server and adds backend capabilities for apps. Installing a plugin does not install server modules. See the repository and compatibility map.
-
Workspace — people with roles (workspace-admin / editor / viewer).
-
App — a globally unique slug, owned by a workspace. Its hosts:
<slug>.<APPS_DOMAIN>(published),<slug>--preview.<APPS_DOMAIN>(the newest version that compiled),<slug>--v<N>.<APPS_DOMAIN>(exactly version N), plus verified custom domains. -
Version — an immutable, numbered snapshot of the app's sources and compiled output. Publishing moves one pointer.
-
Platform modules — the only backend an app gets: routes under
/__drobek/v1/<name>on the app's host,drobek.<name>in the browser SDK, a per-app config the agent proposes and the owner confirms when it is risky, a skill the agent reads. Built in (modules/, enabled withDROBEK_MODULES):auth— the app's users sign in with an e-mailed code (allowlist, admins,<LoginGate>, sessions the owner can revoke);data— collections of records with per-operation rules (public/user/owner/admin), JSON Schema, CSV;forms—<Form>submissions stored and e-mailed to the owners, with bot protection;email— notifications to the app's owners (never to arbitrary addresses), per-app and server-wide budgets;files— end-user uploads with types sniffed from the bytes and a per-app quota;proxy— calls to external APIs with the secret injected server-side, behind an SSRF guard.
Operators can add their own modules against the public contract:
npm create drobek-module@latest <name>scaffolds one against the npm packages@freema/drobek-modules+@freema/drobek-sdk(imported as@drobek/modules/@drobek/sdkthrough npm aliases) —docs/MODULES.md→ Writing a module.
Self-host quickstart
What you need:
-
a server with a public IPv4 (and/or IPv6), linux/amd64 (there is no ARM image in v1), ports 80 and 443 reachable from the internet;
-
a domain for the dashboard and a domain for the apps. Create these DNS records before step 3 (Let's Encrypt checks them):
A separate registrable domain for the apps (
example.netnext toexample.com) is the safer choice;apps.<your dashboard domain>works too. No DNS at all (a test box)? UseDOMAIN=localhostin step 3 — Caddy's local CA (tls internal), reachable only from the machine itself. -
an SMTP account (host, port, user, password, a sender address) or a Resend API key — sign-in codes go out by e-mail.
Every command runs as root (or prefix sudo).
1. Docker, git and go-task
2. The drobek files (the compose file, the scripts, the env template — the image itself comes from GHCR)
3. Configuration — generates every secret, writes .env.production
(mode 600), renders deployments/Caddyfile with the image's own generator
(no Node on the host):
Expected output (abridged):
Then put the SMTP password in (never on the command line):
task selfhost:init never asks anything and never overwrites a secret; run it
again whenever you like (after changing TLS settings: then task tls:reload).
The TLS default for a real domain is on-demand (one Let's Encrypt
certificate per app host, gated by drobek); TLS_MODE=wildcard-file or
TLS_MODE=dns pick a wildcard certificate instead — see "TLS" in
docs/SELF-HOSTING.md.
4. Start
The first start pulls the images and drobek applies every database
migration. Expected: Container drobek-prod-postgres-1 Healthy, …-redis-1 Healthy, …-drobek-1 Healthy, …-caddy-1 Healthy.
Tip: alias dc='docker compose --env-file .env.production -f docker-compose.production.yaml'
— the rest of this guide spells the command out.
5. Sign in — open https://drobek.example.com, enter your
SUPERADMIN_EMAIL, type the 6-digit code from the e-mail. You land on /me
with a personal workspace.
6. Connect an agent (Claude Code)
Claude Code discovers drobek's OAuth server, opens the consent page in your
browser (read, write, publish) and gets a token bound to you. Without a
browser on the agent's machine, mint an API key on the server instead and
pass it as a header:
7. Publish an app — ask the agent: "Build a tip calculator on drobek and
publish it." It calls create_app → write_files → publish; open the
published_url it returns (https://tip-calculator.apps.example.net). With
on-demand TLS the very first request to a new app host waits a few seconds for
its certificate.
A test box without DNS — the same steps with Caddy's local CA:
Trust drobek-root.crt in your browser / OS to use it without warnings (Node
clients: NODE_EXTRA_CA_CERTS=drobek-root.crt). App hosts are
https://<slug>.apps.localhost, which browsers resolve to the machine itself.
HTTPS_PORT=8443 (plus HTTP_PORT=8080) moves Caddy off 443 — every URL then
carries the port.
Backups (task backup / task restore), upgrades (task selfhost:upgrade),
the three TLS paths, custom domains, abuse handling and every setting:
docs/SELF-HOSTING.md.
Connect your agent
The MCP endpoint is <PUBLIC_APP_URL>/mcp (OAuth 2.1; the client discovers,
registers and asks for your consent by itself). Scopes: read, write,
publish; the token is bound to you and reaches every workspace you belong
to, with your role in each.
- Claude Code:
claude mcp add --transport http drobek https://drobek.example.com/mcp, then/mcp→ sign in. For the hosted drobek:claude plugin marketplace add freema/drobek-pluginandclaude plugin install drobek@drobek(MCP server + build skill +/drobek:build-app). - Claude (web / desktop): add a custom connector with the
/mcpURL. - Cursor:
~/.cursor/mcp.json→{ "mcpServers": { "drobek": { "url": "https://drobek.example.com/mcp" } } }. - Codex: for the hosted drobek
codex plugin marketplace add freema/drobek-plugin,codex plugin add drobek@drobek,codex mcp login drobek. - Scripts / CI: a personal
drk_…API key from/me/api-keysasAuthorization: Bearer drk_….
The agent gets twenty-five tools — list_apps, create_app, duplicate_app, get_app,
read_file, write_files, restore_version, publish,
set_gallery_listing, skill_info, configure_module, query_data,
get_logs, for video, audio, images and fonts create_asset_upload,
list_assets, delete_asset (an upload URL — the file never passes through
the model), and for custom domains list_domains, add_domain,
verify_domain, set_primary_domain, remove_domain, and for the proxy
module's external APIs list_upstreams, register_upstream,
remove_upstream, and sync_now for the sync module's scheduled imports (a super-admin also gets set_workspace_publishing). The full agent contract (scopes,
the briefing, skills, /llms.txt) is docs/AGENT.md; a
running server serves it at /llms.txt, /llms-full.txt and
/build-with-your-agent. To teach an agent the loop without the plugin:
cp -r skills/drobek ~/.claude/skills/drobek.
Develop locally
Prereqs: Docker (compose v2) for the stack; go-task
3, Node 22 and pnpm 10 for the host-side task check / task e2e.
The dev stack runs every built-in module plus the example
drobek-module-hello. Sign in at localhost:3041
(the code is in Mailpit), then point your agent at
http://localhost:3041/mcp.
Opening apps locally. APPS_DOMAIN=apps.localhost:3041, and browsers
resolve every *.localhost name to your machine — nothing to add to
/etc/hosts:
curl and Node do not resolve *.localhost everywhere; send the app host in
the Host header instead:
Browsers refuse __Host- cookies on plain http://localhost, so the http dev
stack (NODE_ENV ≠ production) uses the unprefixed, still host-only
drobek_session / drobek_app_access / drobek_eu; production and every
https origin use the __Host- names. task dev:tls runs the dev stack behind
Caddy on https://localhost with its local CA. For scripts without an OAuth
flow: task api-key:create [email protected] NAME=laptop SCOPES=read,write.
Everyday commands:
Contributor rules and the repository map: CLAUDE.md.
Public repositories, agent plugins versus server modules, and the external
consumer check: docs/ECOSYSTEM.md.
End-to-end tests
tests-e2e is one Playwright suite with two tiers:
-
@local— needs a local stack: seeds and reads Postgres / Redis / Mailpit directly.global-setup.tsTRUNCATEs the core tables first, but only withALLOW_DESTRUCTIVE=1AND aDATABASE_URLhost on its allow-list (localhost,127.0.0.1,postgres). -
@smoke— public HTTP + MCP only, safe against production: never the database, Redis or Mailpit. The MCP smoke loop (mcp-loop.spec.ts) signs in with adrk_key fromSMOKE_API_KEY(read from the environment only), writes, previews and publishes asmoke-*app, and leaves nothing behind. MCP has no delete tool, so the spec cleans up by target: against the local stack (TEST_ENV=local) it creates a freshsmoke-<random>app and deletes it at the end — also when the test fails — through the dashboard delete action, signed in as the smoke user by e-mail OTP (Mailpit); against any other target (production) it re-uses ONE stable app per key,smoke-<12 hex of a SHA-256 of the key>(list_apps→get_app, then a new version + publish), so production keeps exactly one smoke app per smoke identity:Without the key it is skipped on a localhost target and fails anywhere else; under
task e2eit mints a throwaway key in the local DB.
mcp-loop.spec.ts is the agent loop end to end: the official MCP SDK client
with an OAuth provider (401 → discovery → DCR → PKCE consent driven by
Playwright → token), then list_apps → create_app → write_files (compile
error → fix) → preview host → publish → production host → restore_version
→ get_app, asserted under 90 s.
task e2e runs against the dev stack (task dev). task e2e:image is the
CI flow (.github/workflows/ci.yml runs the same scripts/e2e-image.sh):
build the production image, start it with docker-compose.e2e.yaml (project
drobek-e2e, its own loopback ports, so it runs next to the dev stack) behind
Caddy with tls internal on https://localhost:8443 and
https://<slug>--preview.apps.localhost:8443, let it migrate a fresh DB, then
run @smoke + @local and tear everything down. DROBEK_IMAGE=… skips the
build, E2E_KEEP=1 keeps the stack, extra args go to Playwright
(task e2e:image -- tests/mcp-loop.spec.ts).
Versions and upgrades
drobek follows semantic versioning; while it is at 0.x, a minor release may
need an operator step, and its release notes say which. Every release is a git tag vX.Y.Z, an
immutable image ghcr.io/freema/drobek:vX.Y.Z, a
GitHub release and an entry in
CHANGELOG.md; latest and previous move with each
release (docs/SELF-HOSTING.md → Image tags).
Migrations only go forward and run as their own step of
task selfhost:upgrade; a rollback is the previous image plus, when the
release migrated the database, the backup taken before it
(Upgrades and rollback).
Modules declare the contract range they need, and the server refuses to start
with a module it cannot satisfy
(docs/MODULES.md → Compatibility).
Extend drobek: write a module
An app's backend is a set of platform modules, and anyone can write one: an npm package against the published contract, installed by the operator without building an image.
The contract is on npm as @freema/drobek-modules; the scaffold installs
it under the name module code imports, @drobek/modules
("@drobek/modules": "npm:@freema/drobek-modules@^X.Y.Z").
docs/MODULES.md → Writing a module
walks through the contract, publishing and installing;
examples/drobek-module-hello is a working
one, and Published modules lists the
modules others can install.
Contributing
Issues, questions and pull requests are welcome — see
CONTRIBUTING.md. Questions and ideas go to
Discussions; security issues
are reported privately (SECURITY.md).
Documentation
License
AGPL-3.0 — see docs/LICENSING.md.
Source: README.md at commit 4c49301
Tools
0Version history
1- v0.6.0LatestSep 29, 2026
