Provider routing built for the way agents actually run.
Blackbox takes your compiled LangGraph workflow and routes every LLM call to the provider that fits your cost, latency, and quota constraints — without changing a line in your agent code.
Routers built for chat APIs miss what agents do.
An agent is a graph: each node has a different cost profile, a different latency budget, and different reliability requirements. Blackbox treats the workflow as a first-class object, not as opaque traffic.
- Compile once. Blackbox understands the workflow graph, not just individual calls.
- Cost, latency, and reliability tradeoffs you can see and select from.
- Quotas enforced at the edge of every provider, per workflow.
- No model lock-in: providers and credentials live in your dashboard, not in your code.
Set api-key to the workflow key minted in the dashboard. Blackbox handles provider selection, retries, and quota arbitration.
// Register the workflow in the dashboard first — copy its Workflow API key.const ALLOCATOR_BASE = "https://your-allocator.example.com"; // not the dashboard URL const res = await fetch(`${ALLOCATOR_BASE}/blackbox/invoke`, { method: "POST", headers: { "content-type": "application/json", "api-key": "<workflow-api-key-from-dashboard>", }, body: JSON.stringify({ "focused_node": "route_planner", "prompt": prompt, }),}); const { provider_id, latency_ms, cost_usd } = await res.json();The controls you need once routing is live.
After Blackbox is in the request path, the dashboard gives teams a clear view of provider tradeoffs, health, quota pressure, and workflow behavior.
Compile each LangGraph workflow once. Blackbox understands the graph and routes per-node so your agent stays correct.
See every viable provider permutation on a cost vs. latency frontier. Pick the tradeoff that matches your budget.
Latency histograms, error rates, cost per call. Observability that sticks to the model and provider, not just the node.
Set absolute RPM/TPM/daily caps per provider per workflow. Blackbox shifts traffic before you hit a 429.
Path-change detection across runs surfaces silent regressions in your compiled workflow before users notice.
One control plane: register workflows, manage providers, change tradeoffs, inspect runs — all in one place.