BrainstormRouter TypeScript SDK

Agent orchestration SDK with multi-turn session loops, per-turn model selection, governed tool execution, and OpenAI-compatible proxy across 31 curated models from 8 providers (Anthropic, OpenAI, Google, xAI, Groq, Perplexity, DeepSeek, Moonshot).

Install

npm install brainstormrouter

Agent Runs (Session Loop)

Run multi-turn agent sessions where Thompson sampling picks the best model per turn, tools execute through a governed MCP pipeline, and budgets enforce in real-time:

import BrainstormRouter from "brainstormrouter";

const client = new BrainstormRouter({ apiKey: "br_live_..." });

// Bootstrap an agent with identity and budget
await client.agentProfiles.bootstrap("research-01", {
  display_name: "Research Agent",
  budget_daily_usd: 10.0,
  department: "Engineering",
});

// Run a multi-turn agent loop — streams SSE events
const result = await client.agentRuns.run("research-01", {
  message: "Research the latest AI safety papers and summarize findings",
  max_turns: 10,
  budget_cap_usd: 2.0,
});

Session Management

// Create a session without running
const session = await client.agentRuns.createSession("research-01");

// Check session status
const status = await client.agentRuns.getSession("research-01", session.id);

// Send a message to a running session (mid-run steering)
await client.agentRuns.sendEvent("research-01", session.id, {
  type: "user.message",
  text: "Also check for papers on alignment",
});

// Interrupt a running session
await client.agentRuns.sendEvent("research-01", session.id, {
  type: "user.interrupt",
});

// View run history
const { runs } = await client.agentRuns.listRuns("research-01", { limit: 10 });

Proxy Mode (OpenAI-Compatible)

Also works as a drop-in replacement for the OpenAI SDK — compatible with LangChain, Vercel AI SDK, CrewAI, and any OpenAI-compatible framework:

const response = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-4",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(response.choices[0].message.content);

Use "auto" as the model to let Thompson sampling pick the best model:

const response = await client.chat.completions.create({
  model: "auto",
  messages: [{ role: "user", content: "Explain quantum computing" }],
});

Streaming

const stream = await client.chat.completions.create({
  model: "auto",
  messages: [{ role: "user", content: "Write a poem" }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

Memory (4-Block RMM)

// Append a memory entry
await client.memory.append("User prefers concise answers", {
  block: "semantic",
});

// List memory entries
const { entries } = await client.memory.entries();

Guardrails

// Test content against guardrail pipeline
const result = await client.guardrails.test({
  content: "Check this message for PII",
  direction: "inbound",
});

Provider Keys (BYOK)

// Register your own provider key
await client.providers.register("anthropic", {
  apiKey: "sk-ant-...",
});

Configuration

const client = new BrainstormRouter({
  apiKey: "br_live_...", // or set BRAINSTORMROUTER_API_KEY env var
  baseURL: "https://api.brainstormrouter.com", // default
  maxEstimatedCost: 0.1, // Guardian: cap cost per request
  guardianOff: false, // disable Guardian Intelligence
});

Scope

This SDK wraps the public /v1/* API (API-key authenticated). The dashboard-only /auth/* bridge routes — onboarding wizard state, tenant settings, insights rollups, forensic snapshot listings, admin-only seeders — are intentionally not surfaced here: they're consumed by the Vite dashboard directly and aren't part of the external API contract.

Resources

• Documentation
• API Reference
• Full API Reference (llms-full.txt)
• GitHub
• Dashboard

Specifications

Open, citable, CC BY 4.0:

• Trust Envelope v1 — EdDSA-signed agent identity JWT schema
• MCP-OAuth Bridge v1 — Trust Envelope ↔ OAuth 2.1 protected resource metadata mapping

License

MIT