Agent Tools and MCP
Tools let an Agent-mode conversation retrieve live information or perform an authorized action. MCP is one way to connect external tool servers; GeniSpace also provides platform, built-in, custom-operator, and local application tools.
Tool sources
| Source | Examples |
|---|---|
| Platform tools | Dataset search, Datasource operations, tasks, storage, applications |
| Assistant tools | Create or manage supported GeniSpace resources |
| Custom operators | Team-defined OpenAPI or operator implementations |
| MCP servers | External tools exposed through Model Context Protocol |
| Local application tools | Workbench draft editing and other host-provided actions |
Configure tools for an agent
- Open Console → Agents and edit a conversational agent.
- Open the tools or MCP configuration.
- Enable only the required sources.
- Choose specific operators, tasks, data sources, datasets, or servers where the interface provides a scope selector.
- Configure separate read/create/update/delete access where available.
- Save and test in Chat using Agent mode.
Tool availability is the intersection of agent configuration, current user permissions, space access, and runtime policy.
Add an external MCP server
- Open the MCP server configuration for the agent or space.
- Enter the server name and URL.
- Add authentication through the supported secure credential fields.
- Test discovery.
- Review the discovered tool names and descriptions before enabling them.
Do not place credentials in prompts, descriptions, URLs, or screenshots. Use HTTPS in production.
How a tool call appears in Chat
The execution timeline shows the selected tool, status, duration, and a result summary. Select the event to inspect the structured result in the Activity Panel. Application plugins may render a specialized preview, table, form, or confirmation instead of raw JSON.
If the result contains no records, the final answer must report that outcome. A completed response does not continue executing tools in the background unless the interface explicitly shows a running job.
User confirmation
An interactive agent can pause before continuing when information is missing or an explicit choice is needed. The confirmation card supports:
- Single choice
- Multiple choice with limits
- A custom text answer
After submitting, the saved turn resumes. For destructive or externally visible operations, verify the target and scope before continuing.
Dataset tool selection
| Request | Preferred tool |
|---|---|
| Exact ID, status, date, or known field value | Structured query |
| Literal keyword or identifier | Full-text search |
| Similar experience, capability, meaning, or cross-language concept | Vector search |
Only request output fields that exist in the Dataset schema. Vector Top-K results are similarity candidates, not proof that a mandatory condition is satisfied.
Troubleshooting
| Problem | Check |
|---|---|
| Tool is not selected | Use Agent mode and make the requested action explicit |
| Tool is missing | Verify agent configuration, user permission, space access, and server discovery |
| Invalid field or filter | Inspect the Dataset or tool schema and use declared field names and expression syntax |
| Repeated identical call | Review the first result; refine the request instead of repeating an unchanged read |
| Result is visible only as a summary | Open the event in the Activity Panel or verify that the corresponding Chat plugin is enabled |