Capabilities
Give an agent specific tools, not unrestricted access to your systems.
Administrators register capabilities and control their department availability. Each integration needs configuration, credentials where applicable, and validation against the target system. These categories are not a catalogue of ready-made vendor integrations.
MCP tools
Expose tools from a configured Model Context Protocol server. Availability, authentication and behavior depend on that server and the registered tools. A tool's read/write classification and approval settings must match its actual effects.
Database queries
Run configured, parameterised queries against PostgreSQL or ClickHouse. This is not unrestricted model-generated SQL. Use read-only database credentials for read capabilities; parameter binding alone does not make a query read-only.
HTTP calls
Call a configured HTTP endpoint with a method, headers and optional body template. Bearer-token and API-key authentication can use stored secret references. Endpoint reachability and outbound URL checks constrain what can be called.
Instruction skills
Add task-specific instructions to the agent's planning context. Skills guide model behavior; they are not standalone integrations and do not grant access to external systems.
Configuration reference
All categories use an action_type, name, category and scope. Approval is enabled by default. Database and HTTP capabilities need a JSON Schema object in params_schema, either at the top level or in their configuration.
| Configuration | Fields and behavior |
|---|---|
| db_config | db_type (postgresql or clickhouse), host, port, database, username, password and query_template. The password is encrypted server-side. Templates use named placeholders such as {days}, bound as query parameters, not template-language expressions. |
| http_config | url, method (GET by default), optional headers, auth_type, auth_secret, body_template and read_only. The executor supports bearer and api_key authentication; api_key uses X-API-Key. Body placeholders such as {name} receive JSON-encoded parameter values. |
| mcp_config | Provide server_id for an existing server or server to register one inline. Inline server configuration includes name, base_url and optional authentication fields. Individual tool settings belong to the registered MCP tools. |
| skill_config | instructions supplies natural-language guidance to the planner. Skills have no tool parameters or external execution of their own. |
HTTP calls that send a body or use a non-GET method require approval unless explicitly marked read_only. That flag is an administrator assertion, not verification of endpoint behavior. No universal database-version or third-party service compatibility is implied.
Approval is a workflow, not a sandbox
Capabilities require approval by default. Permitted read capabilities can run directly; write actions use pending approvals. Review scope, parameters, credentials and downstream side effects before enabling a tool. Approval does not guarantee a safe result or an available rollback.
For request schemas, use the API reference generated by your deployed version.Administration guide