Agent Control
Tool Registry
Tool Registry is your organization's catalog of approved tools: MCP servers, HTTP APIs and the knowledge catalogue are defined here. It decides which project can use which tool, sends calls that need approval to a person and logs every call. Your application can also pass its own tools through the OpenAI-compatible endpoint.
Tool calls
The chat API accepts up to 128 tools plus the tool_choice and parallel_tool_calls fields, and forwards them to the provider. Your application runs the tool. You turn it on per project with tools_enabled; it is off by default.
Scope
- Catalog
role 1: Which tools exist, what do they do, who owns them? - Gatekeeper
role 2: Can this project use this tool? Is approval required? - Ledger
role 3: Who ran which tool, when, and what was the result?
| Connection type | What it connects | Rule |
|---|---|---|
| HTTP API | Your existing internal services | URL fixed in the registry; allowed hosts list |
| MCP server | Standard tool exposure | Hidden until an admin approves |
| Knowledge catalogue | Search across your documents | Read-only |
The registry is a permission and control layer; the tools that do the work are your organization's own systems. It doesn't download tools from the internet, and it doesn't replace your application's own user authorization.
The Tool Registry is on in every license. In the Individual license the number of MCP and HTTP connections and of knowledge catalogues is capped; the limits are shown on the app's License screen.
Risk and approval
| RISK | Definition | What happens |
|---|---|---|
| Read | Reads, doesn't change anything | Usually runs without approval, logged |
| Write | Creates or changes records | Human approval when "needs approval" is on |
| Destructive | Deletes or does something irreversible | Human approval when "needs approval" is on |
A project must be granted permission to use a tool. If an unpermitted tool is requested, the call is stopped and logged in Audit Log as "blocked". Audit Log records events, not content: project, tool, decision, result and token counts.
Security rules
- Tool output is labeled as data; it can't change permissions
- A tool that needs approval always goes to approval, whatever the model says
- Each tool can have a timeout and a per-minute call limit
- Connection secrets are stored with AES-256-GCM under a separate key
- The model can call tools at most 5 times per response