Capabilities

Capabilities are the product areas and runtime primitives that users interact with when building and operating SOIT workspaces. Every capability page follows the same shape: what it is, when to use it, how it connects to other capabilities, and which API and UI surfaces support it.

Primary capabilities

• Agents: versioned agent definitions and releases — the primary business object.
• Workflows: graph-based automation with governed node execution.
• Knowledge: document ingestion, chunking, embedding, and retrieval with citation evidence.
• Models: model providers, catalog entries, and routing choices.
• Plugins: packaged capability extensions and the public lifecycle boundary for tools, MCP servers, and skills.

Calling SOIT from your own code

• OpenAI-Compatible Gateway: OpenAI SDKs and frameworks call workspace models under /v1, and the same keys call tools directly — every call a governed run. From v1.4.0 the gateway can run in processes of its own.
• MCP Server: Claude Code, Cursor, and other MCP clients call the workspace's governed tools at /mcp.
• SOIT CLI: run agents, replay regression sets on a new model, and export evidence from a terminal or CI.

Runtime and governance capabilities

• MCP and Skills: plugin-owned MCP servers and skill artifacts.
• Chat and Responses: threaded conversational execution and response projections.
• Tasks and Runs: the execution ledger and async work tracking.
• Approvals: tool calls held for a human decision, with approvers, deadlines, delegation, and history.
• Evidence and Exports: run evidence bundles and ledger exports in a versioned contract.
• Memory: scoped, governed agent memory.
• Evaluation: repeatable quality runs, release gates, and model replays.
• Notifications: runtime and workspace event delivery.
• Identity and Access: users, workspaces, roles, and scoped permissions.
• Secrets and Security: secret references, egress policy, private networks for model servers on your own network, and governed gateways.
• Budgets and Cost: cost attribution, daily usage aggregates, and budgets that alert or stop spend.

Operating a deployment

• Editions and Licenses: Community without a license, Enterprise with a signed one, and extension packages (new in v1.4.0).
• Telemetry: opt-in anonymous usage reports, off by default (new in v1.4.0).
• Configuration: the deployment settings, including the process role and private networks.

How capabilities compose

The capabilities are not silos — they compose through typed bindings and the shared runtime:

• An agent binds models, knowledge, tools, workflows, and memory into one versioned unit.
• Plugins are how tools, MCP servers, skills, and workflow nodes enter a workspace.
• Every execution — chat turn, workflow run, evaluation, gateway call, direct or MCP tool call — lands in the same runs ledger.
• Identity, secrets, and security policy apply uniformly across all of the above.

If you are new to SOIT, start with Agents and follow the links outward.