Skip to content
trainingOSby JamBox
AccessRegister free 中文

Agent & integrations

Give agents a governed path to real work.

Xiaojiang answers from versioned knowledge. MCP tools follow identity, permissions, preview, confirmation and audit.

Discuss integration
A training venue team coordinating daily operations
AI and external Agent access are excluded from the basic version

The following describes governed capabilities. Enabling model processing or external delivery requires provider checks and the applicable notice and consent.

Two agent experiences, one safety boundary.

Xiaojiang support

Helps users understand the current product and move to a human advisor when the request needs business judgement.

  • Answers from published, version-matched help content
  • Shows citations and an explicit provider status
  • Does not read product feedback or customer business records
  • Stops automatic replies after human handoff

Operational MCP

Lets an authorized employee discover and invoke scoped tools without bypassing TrainingOS business services.

  • OAuth 2.1 with PKCE for public authorization
  • Connection, employee, location and tool scopes stay bound
  • Higher-risk writes require preview and confirmation
  • Audit protects sensitive fields and preserves necessary evidence

The model never becomes the permission system.

Authenticate

Resolve the real employee and authorized connection.

Scope

Apply location, permission, field and object-state rules.

Preview

Return blockers and expected changes before a write.

Confirm

Bind the final action to the approved parameters and version.

Audit

Record a redacted, replay-safe business result.

Current standard tool contract.

Tools are grouped into query, preview and confirmed execution. Actual discovery and access depend on the account, plan and scope.

01

Read tools

Query business context without changing state.

02

Preview tools

Validate a proposed action and return blockers.

03

Confirm tools

Apply a pre-approved low-risk operation once.

Plan an integration around one bounded action.

Start with the actor, location, target state, confirmation rule and audit evidence.

Discuss integration