Tool loop

The client sends an ordinary chat request with no tools. The gateway adds the granted servers' tools, runs each tool call the model makes against its MCP server, appends the result, and asks the model again until it answers. Registering servers and calling them directly are covered in MCP gateway.

Turn it on

1. 1

   Register the MCP server

   Add the server on the MCP Servers page. The page checks each server when it opens and lists its tools, so you can see it answers before granting it.

2. 2

   Turn the Tool loop on

   Open Settings › Boons, switch Tool loop on and Open its row. Set Turns, Tool timeout and Request deadline under Limits, and edit the Tool nudge if you need to. Save with the save bar.

3. 3

   Grant the server to the model

   On the model's page, turn on Function calling and Tool choice under Native; the dashboard lets you grant tools only when both are on. Then tick the server under Tool servers and save. You can also grant it from the server's panel on the MCP Servers page.

Or through the Management API:

curl -X PUT http://localhost:9180/api/v1/settings/boons \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "tool_loop_enabled": true,
    "tool_loop_max_turns": 4,
    "tool_loop_tool_timeout_ms": 30000,
    "tool_loop_deadline_secs": 300
  }'

curl -X PUT http://localhost:9180/api/v1/models/$MODEL_ID \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "upstream_model": "Qwen/Qwen3-8B",
    "api_base": "http://inference.example.internal:8000/v1",
    "supports_function_calling": true,
    "tool_servers": ["searxng"]
  }'

Ask something that needs a tool

With a web-search server granted, a plain question is enough:

curl http://localhost/v1/chat/completions \
  -H "Authorization: Bearer $OBLETH_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5-3",
    "stream": true,
    "messages": [{"role": "user", "content": "What is the latest vLLM release?"}]
  }'

The response carries x-obleth-boons: tool_loop and contains the final answer only, never tool_calls. The final answer streams token by token. While tools run between turns, the stream sends a one-line marker naming each tool call on the reasoning channel (reasoning_content), not in the answer.

What the model sees

• The granted servers' tools, merged into the request's tools.
• For a client that sent no tools of its own, the Tool nudge as a system message: it tells the model it has tools and to use them for current or external facts rather than guessing.
• After each tool call, the result as a tool message, cut to 16,000 characters.

A client that brings its own tools, such as an IDE assistant, keeps them. The granted tools are added alongside, the gateway runs only its own, and any call to the client's tools goes back to the client unchanged. No nudge is added for such clients.

Settings

A 0 for any number keeps the stored value.