Playground
Loading connection settings…

Agentic

Tool loop, sandboxes, and MCP integrations. Skills & guardrails are on Policies.

Agentic Gateway

When enabled (or when Global Wiki is on), the gateway may run a multi-step tool loop on the same /v1/chat/completions (or deprecated /api/chat) request. Clients still receive a normal final message.

Turns on ACA/MCP tools for this app. Global Wiki can still work independently.
Tool-calling prompt style. auto picks based on the backend.
How answers are checked before returning to the client.
Whether tools may reach the public internet (restricted = allow-list only).
Hard cap on tool/LLM rounds per user turn.
Whole-loop deadline in seconds. 0 = API default.
Reject empty/short finals under this character count (0 = off).
Comma-separated keywords that pause for human confirmation before tool run.
Comma-separated hosts allowed when egress is restricted.
Optional regexes the final answer should match (comma-separated).
Fail the loop if a shell/python tool returns a non-zero exit code.
Escalate instead of returning a partial answer when the iteration cap is hit.
Execution tools
Code runners for shell/python/node. Use self-hosted-sandbox with http://sandbox-runtime:8080 in Docker, or aca-session with an Azure Dynamic Sessions pool URL. (Different from MCP runtime.)
Language runtime inside the sandbox/pool.
Base URL of the sandbox service (POST /execute).
Integration tools (MCP)
Model Context Protocol servers this tenant may call. Supports http (remote URL) and stdio (local process, Cursor-style).
Example MCP JSON (stdio + http) 
{
  "type": "mcp",
  "name": "example-stdio",
  "transport": "stdio",
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-everything"],
  "env": { "API_BASE_URL": "https://example.com" },
  "credentialRef": "example-stdio",
  "enabled": true,
  "timeoutSeconds": 300
}