We Now Have Built-In MCP for Users. Bring Your Own Agent!

Otto do it has always treated AI as part of the product architecture, not as a chat window bolted onto the side. The native Otto AI layer still matters: it understands projects, tasks, drafts, labels, Activity, and the safety boundaries around product changes. It can inspect work, summarize projects, break down ideas, create reviewed drafts, and prepare task or stream changes through preview and confirmation flows.

That native layer is designed to use stronger model reasoning when reasoning is valuable, while routing ordinary product actions through Otto do it APIs. If a task can be created, moved, commented on, summarized, or read back through deterministic product infrastructure, Otto does not need to spend tokens pretending the model should do everything. The result is a cleaner system: AI for judgment and language, product APIs for reliable execution.

Otto do it also supports a second path: bring your own agent.

[block:mcp-workspace-architecture-diagram]

For technical users who already live with Claude projects, local agents, OpenClaw sessions, or dashboard-based automation, Otto do it can expose one workspace through MCP. MCP stands for Model Context Protocol. In practice, it gives an AI client a structured way to discover tools, request context, and take authorized actions without scraping a screen or asking the user to copy state between systems.

Inside Otto do it, the connection is workspace-scoped. A user enables MCP from Workspace Settings, chooses the client target, and receives an endpoint and credential for that workspace. The connected agent does not become a super admin. It does not get billing powers, account deletion powers, secret access, member-management powers, or a shortcut around product permissions. It receives a bounded tool surface for work that the connected workspace is allowed to expose.

[block:mcp-workspace-settings-assistant]

That tool surface has grown a lot since the first MCP pass. A connected agent can now read workspace context, list projects, inspect task details, search work, read task comments, list workspace members, list project streams, read filtered Activity, and work with labels. It can also use task context packs, read-only resources, reusable prompts, and a non-mutating validate_task_update helper before attempting a write.

The write tools remain intentionally narrow. A connected agent can create tasks, update supported task fields, assign owners, manage labels, attach streams, set due dates and due times, and add comments when the granted scope allows it. Stale-write checks help prevent an agent from overwriting newer human work, and every meaningful write returns readback so the client can cite the task key, changed fields, and Activity context instead of guessing.

There is now a dedicated MCP docs area that expands on how OpenClaw and Claude agents work with Otto do it MCP. It includes setup notes, starting prompts, tool examples, request shapes, response examples, and the limits that keep the connection bounded.

[block:built-in-mcp-tool-examples]

This matters because useful agent integration is not just "let an AI see your app." It is identity, permission, readback, audit, and recovery. It is knowing which workspace the agent can touch, which tool ran, which credential was used, what changed, and where the product can show the result afterward.

That foundation is why MCP is not an experiment sitting outside the product. It uses the same architecture that powers Otto AI and the rest of the app. Native Otto AI remains the integrated assistant for users who want the product to handle the AI experience directly. MCP is for users who already have their own agent stack and want that agent to work with Otto do it as a real tool, not as a browser tab it has to guess its way through.

This is where a lot of software is heading. Larger platforms are only beginning to turn agent access into something that feels operationally serious instead of demo-driven. Otto do it is already building the pieces that make it useful in real work: structured tools, scoped access, clear attribution, safe writes, reusable context, and a product model that can support both native AI and external agents without making either one feel like an afterthought.

Private beta is coming soon. If you want to claim a workspace and try Otto do it with native Otto AI, MCP, or your own connected agent, join the early invite list.