The tools an agent sees depend on which toggles are enabled on your server configuration. For guidance on choosing between direct mode, agent mode, and skill tools, see Choose your pattern.
| Tool | Available when | Purpose |
|---|
| Per-action tools | Agent Mode OFF (default) | One tool per app action or workflow |
RESOLVE_ACTIONS | Agent Mode ON | Resolves user intent to a list of specific actions |
EXECUTE_ACTION | Agent Mode ON | Executes one action or workflow from a resolved plan |
GET_KNOWLEDGE_INDEX | Retrieve Skill ON | Returns a server-scoped index of the skills available to load |
LOAD_SKILL | Retrieve Skill ON | Loads the full content of a specific skill |
Direct mode
When Agent Mode is off, each action and workflow on the server becomes its own tool. Names are auto-generated from the app slug and the action or workflow name.
Example tool names:
salesforce_create_contact_action
salesforce_list_deals_action
sap_get_purchase_order_action
hubspot_sync_leads_workflow
input_payload).
Agent mode
When Agent Mode is on, per-action tools are replaced with two meta-tools: RESOLVE_ACTIONS and EXECUTE_ACTION.
RESOLVE_ACTIONS
Takes the user’s request and resolves it to a list of specific actions to run.
| Parameter | Type | Required | Description |
|---|
integration_query | list[string] | Yes | Action descriptions extracted from the user’s request. Each entry describes one action. |
integration_query: Break the user’s request into individual actions. Each entry should be a short description of the action and the target app.
Example:
User: "Get all my Salesforce contacts and create a note in HubSpot"
integration_query: [
"List Contacts Salesforce",
"Create Note HubSpot"
]
[
{
"slug": "salesforce",
"type": "action",
"identifier": "get_contacts",
"json_schema": {
"type": "object",
"properties": {
"where": {"type": "string"},
"order_by": {"type": "string"}
}
}
},
{
"slug": "hubspot",
"type": "action",
"identifier": "create_note",
"json_schema": { "...": "..." }
}
]
| Field | Type | Description |
|---|
slug | string | Application slug (for example, salesforce or hubspot). Pass this as application_slug to EXECUTE_ACTION. |
type | "action" | "workflow" | Whether the entity is an action or a workflow |
identifier | string | The action or workflow ID. Pass this as action_id to EXECUTE_ACTION |
json_schema | object | JSON Schema for the action’s input. Pass through to input_payload. May be {} if the entity has no inputs. |
When nothing matches the request, RESOLVE_ACTIONS returns a plain string (for example, "Couldn't find any matching actions...") instead of a list. Read the response as text before assuming a plan came back.
EXECUTE_ACTION
Runs a single action or workflow. Call once per resolved action from the RESOLVE_ACTIONS output.
| Parameter | Type | Required | Description |
|---|
application_slug | string | Yes | The app to run against (for example, "salesforce") |
action_id | string | Yes | The action or workflow ID from the resolved plan |
input_payload | object | string | Yes | Input fields for the action, as a JSON object or a JSON string (for example, {"email": "user@example.com"}) |
type | "action" | "workflow" | Yes | Whether this is an action or a workflow |
{
"application_slug": "hubspot",
"action_id": "create_contact",
"input_payload": {
"email": "jane@example.com",
"firstname": "Jane",
"lastname": "Doe"
},
"type": "action"
}
{
"success": true,
"data": {
"id": "contact_12345",
"...": "downstream response from HubSpot, structure depends on the action"
}
}
data is the third-party app’s response passed through, so it varies by action. On failure, the agent receives an error message describing what went wrong, and the same detail is recorded in the audit log.
Available when Retrieve Skill is enabled. Skill tools can be combined with either direct or agent mode.
GET_KNOWLEDGE_INDEX
Returns a compact index of the skills available on this server. Despite the tool name, the “knowledge index” and the skill index are the same thing. The agent browses the index, then loads the full content of a specific entry with LOAD_SKILL.
| Parameter | Type | Default | Description |
|---|
limit | int | 50 | Maximum number of index entries to return |
offset | int | 0 | Number of entries to skip, for paging through a long index |
LOAD_SKILL.
LOAD_SKILL
Loads the full content of a skill by ID or name. At least one parameter must be provided.
| Parameter | Type | Description |
|---|
skill_id | string | The skill’s ID (from GET_KNOWLEDGE_INDEX results or the Skills tab) |
skill_name | string | Look up a skill by name instead of ID |
{
"success": true,
"results": [
{
"file_id": "<skill_id>",
"snippet": "## Goal\nQualify an inbound lead...",
"skill_name": "Lead Qualification",
"category": "Sales",
"apps": ["salesforce"]
}
]
}
results[0].snippet. Content over 100,000 characters is truncated with an explicit marker appended.
LOAD_SKILL never throws. If a skill can’t be loaded, the call still returns cleanly with {"success": false, "error": "..."}, so check the success field in the response before using the content.
Re-authorization
If a connection needs re-authorization (for example, an app’s credentials have expired), the end user is prompted to reconnect through Refold’s hosted flow. The agent does not handle or store credentials itself.
For the underlying token and access model, see Authentication.
Execution sequence
When both Agent Mode and Retrieve Skill are enabled, call the tools in this order:
The tool descriptions instruct agents to follow this ordering and to call EXECUTE_ACTION only after RESOLVE_ACTIONS returns a plan. The ordering is guided by these descriptions, not enforced by a runtime check.
See also
- Architecture: how tools map to actions, workflows, and skills
- Choose your pattern: when to use direct mode, agent mode, or skill tools
- Skills: how agents discover and run multi-step procedures
- Authentication: the token model and access toggles
- MCP Logs: per-call audit trail with input and output
