Doop puts the agent on the canvas, not behind a prompt box
Doop is an open-source design workspace built around a specific idea: people and software agents should edit the same canvas and see the same state as work happens. Each Canvas has shareable Frames that render real HTML inside sandboxed iframes. The project says collaborators can see cursors, presence, edits, agent status, and an activity feed in one WebSocket room. At v0.4.0, this is more concrete than a chat panel that generates a static mockup and asks the designer to refresh.
The built-in MCP server is the most distinctive piece. A Claude Code user can connect with 1 command, approve the standard OAuth flow, and then ask the agent to work on a particular canvas. The agent acts with the permissions of the approving user, while the browser shows its progress. That model makes the output visible and attributable. It also makes Doop relevant to teams already using MCP clients, rather than forcing them into a proprietary prompt interface.
Our clean install, build, and test run all completed
We cloned commit b187ba9 into a fresh, unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. Installation succeeded in 38 seconds. Bun installed 1,867 packages and used 569 MB on disk, which is a sizable dependency footprint for a repository measuring 294 files, roughly 40,621 source lines, and 4 MB checked out. The quickstart was not merely plausible on paper, it worked on our box.
The build then succeeded in 13 seconds. Vitest completed in 68 seconds with 350 tests passing, 0 failing, and 11 skipped out of 361. We also found 5 CI workflow files, a Dockerfile, a compose file, and a tests directory. Those are useful production signals, but skipped tests are not passed tests. A team evaluating Doop should inspect what those 11 cases cover before mapping our clean result to its own authentication, WebSocket, storage, and provider configuration.
The local path is unusually considerate
The README offers 2 credible starting paths. For development, bun run dev starts the web app on port 4300 and the API, WebSocket, and MCP server on port 4400, with PGlite persisting data under data/pg. For a production-style start, Docker Compose runs the application with Postgres on port 4400 after an operator supplies BETTER_AUTH_SECRET. Optional SMTP, stock-photo, object-storage, and analytics integrations are described as graceful additions rather than prerequisites.
That separation gives evaluators a fast route without pretending that an embedded database is the final answer for every deployment. The non-Docker production path, bun run build && bun run start, serves the application from 1 server and accepts DATABASE_URL for Postgres. Still, our 569 MB install is heavy, and the real deployment surface extends beyond whether the process boots. Operators must plan authentication secrets, persistence, backups, reverse proxying, email delivery, and any external integrations they enable.
Live collaboration is the strongest reason to choose it
Doop's feature list hangs together around collaborative work rather than disconnected AI tricks. Frames are HTML artboards, comments can be pinned to elements, undo and redo are shared workspace concerns, and agents expose working status while streaming changes. The design-memory layer adds another useful loop: users can pin exemplar frames, record decisions, and let a distiller propose lasting style rules. The first-canvas performance is honestly labeled as a scripted replay in server/demo.ts, which replays 1 pre-authored frame rather than presenting it as autonomous intelligence.
The built-in design team can pick up a queued card, respond to an @mention, or process task feedback. Roles cover building plus focused passes such as UX, copy, brand, and accessibility, and a card can pass through several roles in order. The server-funded path uses Anthropic by default, with Azure OpenAI available as a server-level provider. After the configured free-task limit, users can connect either a ChatGPT subscription or an OpenAI API key and select among 3 documented tiers: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna.
Credentials, licensing, and youth are real tradeoffs
AGPL-3.0 is another deliberate constraint, not a footnote. It can be a good fit for organizations comfortable contributing changes under that license, but legal review is sensible before modifying and serving the software. Maturity also needs perspective: the repository had 668 stars and 17 open issues in the supplied snapshot, its latest release was v0.4.0 on September 10, 2026, and its last push was September 12. Those dates indicate active development, while the low version number and recent creation signal a product still moving quickly.
Community momentum is promising, not yet conclusive
Fresh code activity and a release only 3 days before this review are healthier signals than an old tag viewed alone. The repository exposes contribution guidance, CI badges, and a Discord link, and the measured 5-workflow CI setup suggests maintenance discipline. However, 17 open issues do not tell us response time, maintainer capacity, or how quickly regressions are resolved. The available community notes include a Product Hunt discussion and a velocity snapshot, but not enough detailed user reports to support claims about broad production adoption.
It belongs beside design and development workflows
Doop fits best as the shared visual surface between a product team and its coding agents. The browser is where people review frames, comments, presence, and decisions; MCP is how an external agent participates; Postgres is the durable store for a serious deployment; and model providers power the optional resident team. It does not replace source control, CI, product analytics, or a full delivery pipeline. A sensible 1-team pilot would test access inheritance, concurrent edits, agent attribution, recovery, backups, and export or handoff expectations before expanding access.
Choose Doop when live, visible agent work is better than copying generated markup between tools. Choose Penpot for conventional design, Excalidraw for collaborative sketching, or tldraw for an embeddable canvas. Our successful 119-second install, build, and test sequence earns Doop a serious pilot.