Server tools

Normally your agent runs the tool-calling loop: the model asks to call a tool, your harness executes it, appends the result, and calls the model again. Server tools move that loop into BitRouter. You declare a set of tools, BitRouter advertises them to the model, and when the model calls one, BitRouter executes it and feeds the result back itself — looping until the model stops calling them. To the caller it looks like a single response.

How the loop runs

BitRouter injects the declared tools into the outbound request, intercepts the model's calls to them, runs them, appends the results, and re-calls the upstream — repeating until the model returns an answer with no tool calls, or a bound is hit. The loop is bounded so it can't run away:

Before each call, an approval policy decides whether the tool may run. The default allows everything; a denied call returns an execution-denied result to the model instead of running.

Enabling server tools per request

You turn server tools on by declaring them in the request's tools array — no config change required. BitRouter recognizes three built-in, model-backed tools and only advertises the ones you declare:

{
  "tools": [
    { "type": "bitrouter:advisor",  "args": { "model": "anthropic/claude-opus-4.8", "instructions": "..." } },
    { "type": "bitrouter:subagent", "args": { "model": "openai/gpt-4o-mini", "instructions": "..." } },
    { "type": "bitrouter:fusion",   "args": { "panel": [{ "model": "..." }], "judge": { "model": "..." } } }
  ]
}

MCP-server tools are wired through configuration instead — set server_tools.mcp_servers to the servers whose tools BitRouter should run inside the loop.

Advisor

Advisor lets the running model consult a stronger model mid-generation. The advisor model is fixed by your declaration (and falls back to the parent model); the calling model sends a prompt and gets back structured advice. Use it when one hard sub-question is worth a brief escalation, without switching the whole request to a pricier model.

SubAgent

SubAgent lets the running model delegate a self-contained task to a cheaper, faster worker model. The worker is fixed by the declaration; the caller supplies a task_name and task_description and gets back the outcome. Use it to fan out bounded sub-tasks without spending frontier tokens on them.

Fusion

Fusion runs a panel of models (1–8) on the same prompt in parallel, then a judge model compares — not merges — their answers into a structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots), which the calling model uses to write the final answer. An optional synthesizer can write that answer instead. Use it for high-stakes questions where cross-checking several models is worth the cost.