Studio, MCP, SDK, and dashboard
Sapiom has several surfaces because local project authoring, direct capability calls, and cloud operations have different ownership boundaries. They share the same agents and capabilities, but they are not interchangeable clients.
Responsibility at a glance
Section titled “Responsibility at a glance”| Surface | Primary responsibility | Runs where |
|---|---|---|
| Agent Studio | Local IDE for editing agent projects with a coding agent; adds Canvas, Steps, Code, and direct lifecycle actions. | Your machine |
Local authoring MCP (@sapiom/mcp) | Scaffold, check, test, link, deploy, run, inspect, signal, and schedule from a local checkout. | Your MCP client’s machine |
| Hosted MCP | Direct capability calls plus cloud agent lifecycle tools for remote clients. | Sapiom cloud |
@sapiom/agent | TypeScript contract for declaring an agent, its steps, graph, inputs, and directives. | Your project and the agent runtime |
@sapiom/tools | Typed capability client, either runtime-injected as ctx.sapiom or authenticated explicitly for standalone calls. | Your code or an agent run |
| Agents dashboard | Browse templates and operate organization-owned agents, versions, runs, alerts, and settings. | Browser and Sapiom cloud |
Agent Studio
Section titled “Agent Studio”Agent Studio hosts Claude Code in a real local terminal and gives each session Sapiom-specific context. It wires both MCP surfaces into the session:
- the local authoring MCP for project and lifecycle work;
- the hosted MCP for direct Sapiom tools.
Studio’s Local Run, Deploy, and Prod Run controls call the local lifecycle implementation directly. Clicking one does not ask the coding agent to type or execute the operation, so it does not consume coding-agent turns.
The two MCP surfaces
Section titled “The two MCP surfaces”The package @sapiom/mcp is the local authoring MCP. Its server identifies itself as sapiom-dev, and its authoring tools use the sapiom_dev_* namespace. The name you choose when registering it in an MCP client is only a client-local alias; it does not rename the server or its tools.
The hosted MCP is a remote, authenticated surface. It exposes direct sapiom_* capability tools and implemented sapiom_workflow_* cloud operations. It does not edit a local project, and its served catalog intentionally excludes some internal or unsupported tools. Use Hosted MCP direct access when you need this surface.
The TypeScript packages
Section titled “The TypeScript packages”“SDK” covers two distinct public contracts:
@sapiom/agentdefines the automation: steps, transitions, input contracts, and control-flow directives.@sapiom/toolscalls Sapiom capabilities. Inside a production step, use the injectedctx.sapiom; in standalone code, create an authenticated client.
The distinction is useful: authoring an agent graph is not the same operation as calling one capability directly. See the SDK and API reference for direct access and the authoring guide for agent code.
Agents dashboard
Section titled “Agents dashboard”The Agents dashboard is the cloud fleet and operations surface. It can materialize a template, start a run, inspect results, and configure an owned agent. It does not currently edit the project’s source code, and it does not provide schedule management.
Use Agent Studio or Claude Code with the local authoring MCP to change source. Use the dashboard once you need organization-wide visibility into hosted agents and production runs.
© 2026 Sapiom, Inc.