Bridges

LLM connectors. Pick one, swap any time.

Bridges

A bridge is the thing the body calls when local layers cannot answer. All bridges extend SCPBridge and implement call(prompt, tools).

Bridge	Models	Cost (per 1k tokens)	Setup
OllamaBridge	llama3.2, mistral, qwen, phi	$0.000	Install Ollama, pull model
BedrockBridge	Nova Micro, Claude via Bedrock	$0.00013 (Nova Micro)	AWS credentials
OpenAIBridge	gpt-4o-mini, gpt-4o	$0.00015 (mini)	OPENAI_API_KEY

OllamaBridge

Free, local, no API key. Default for getting started.

ollama pull llama3.2
ollama serve   # starts on http://localhost:11434

const { OllamaBridge } = require("scp-protocol/bridges/ollama")

const bridge = new OllamaBridge({
  model: "llama3.2",
  host: "http://localhost:11434",
  systemPrompt: "you control a cart-pole",
  maxTokens: 256,
  temperature: 0.1,
})

const result = await bridge.call({ entity: "drone" })

OllamaBridge.isAvailable() is a quick health check that returns a boolean and never throws.

BedrockBridge

Wraps the AWS SDK Converse API. @aws-sdk/client-bedrock-runtime is an optional peer dep.

npm install @aws-sdk/client-bedrock-runtime

const { BedrockBridge } = require("scp-protocol/bridges/bedrock")

const bridge = new BedrockBridge({
  model: "amazon.nova-micro-v1:0",
  region: "us-east-1",
  systemPrompt: "...",
})

Credentials follow the standard AWS chain (env, shared file, EC2 role).

OpenAIBridge

Raw node:https. No SDK dependency.

const { OpenAIBridge } = require("scp-protocol/bridges/openai")

const bridge = new OpenAIBridge({
  model: "gpt-4o-mini",
  apiKey: process.env.OPENAI_API_KEY,
})

Writing a custom bridge

Subclass SCPBridge. Implement call(prompt, tools). Return whatever shape your body expects.

const { SCPBridge } = require("scp-protocol")
const https = require("node:https")

class MyBridge extends SCPBridge {
  constructor(opts = {}) {
    super(opts)
    this.endpoint = opts.endpoint
    this.token = opts.token
  }

  async call(prompt, tools) {
    const body = JSON.stringify({ prompt, tools, model: this.model })
    const raw = await this._post(body)
    const parsed = JSON.parse(raw)
    return { decision: parsed.choice, raw: parsed }
  }

  _post(body) {
    return new Promise((resolve, reject) => {
      const req = https.request({
        method: "POST",
        hostname: this.endpoint,
        port: 443,
        path: "/v1/decide",
        headers: {
          "Content-Type": "application/json",
          "Authorization": `Bearer ${this.token}`,
        },
      }, (res) => {
        let data = ""
        res.on("data", (c) => { data += c })
        res.on("end", () => resolve(data))
      })
      req.on("error", reject)
      req.write(body)
      req.end()
    })
  }
}

The base class tracks callCount, errorCount, totalDurationMs, and lastCallMs. Read them via bridge.stats().

Cost comparison

Sample run: 100 brain calls during a 10-minute session, ~400 tokens per call.

Bridge	Total cost
OllamaBridge	$0.00
Nova Micro	$0.0052
GPT-4o Mini	$0.0060
Claude Haiku	$0.0100

After the cache learns the typical situations, brain calls drop to a handful per minute and the numbers above shrink proportionally.