Skip to content

Cloudflare Integration Overview

Bernstein can run agents locally or in the cloud. The Cloudflare integration lets you execute agents on Cloudflare's edge infrastructure using Workers, Durable Objects, Workflows, and V8 isolate sandboxes -- while the orchestrator stays deterministic and local.

When to use cloud vs local

Scenario Recommended Why
Solo developer, small codebase Local Zero setup, fastest iteration
CI/CD pipeline Local (Docker/K8s) Full filesystem access, no network hops
Team with shared orchestration Cloudflare Centralized billing, no shared server to maintain
Untrusted code execution Cloudflare sandbox V8 isolate/container isolation, no host access
Global team, low-latency agent dispatch Cloudflare Workers Edge execution near developers
SaaS / hosted Bernstein Cloudflare (full stack) D1 analytics, R2 storage, Vectorize caching

Architecture

graph TD
    User["Developer / CI"]
    CLI["bernstein cloud CLI"]
    Orch["Local Orchestrator<br/>(deterministic tick loop)"]
    Bridge["RuntimeBridge<br/>(cloudflare / cloudflare-workflow / cloudflare-sandbox)"]

    subgraph Cloudflare["Cloudflare Edge"]
        Worker["Workers + Durable Objects<br/>(agent lifecycle)"]
        Workflow["Workflows<br/>(durable multi-step execution)"]
        Sandbox["V8 Isolate / Container<br/>(code sandbox)"]
        R2["R2 Object Storage<br/>(workspace sync)"]
        D1["D1 (SQLite)<br/>(analytics & billing)"]
        Vec["Vectorize<br/>(semantic cache)"]
        AI["Workers AI<br/>(free LLM provider)"]
        Browser["Browser Rendering<br/>(web browsing)"]
    end

    MCP["MCP Remote Transport<br/>(streamable HTTP)"]

    User --> CLI --> Orch
    Orch --> Bridge
    Bridge --> Worker & Workflow & Sandbox
    Worker --> R2
    Sandbox --> R2
    Workflow --> Worker
    Orch --> AI
    Orch --> Vec
    Orch --> D1
    Orch --> Browser
    MCP --> Orch

Module map

Module Import path Purpose
Workers RuntimeBridge bernstein.bridges.cloudflare Spawn agents on Cloudflare Workers with Durable Objects
Workflow Bridge bernstein.bridges.cloudflare_workflow Durable multi-step workflows with auto-retry and approval gates
Sandbox Bridge bernstein.bridges.cloudflare_sandbox Isolated V8 or container execution for untrusted code
Browser Rendering bernstein.bridges.browser_rendering Headless browsing, screenshots, scraping, PDF generation
R2 Workspace Sync bernstein.bridges.r2_sync Content-addressed file sync between local and R2
Workers AI Provider bernstein.core.routing.cloudflare_ai Free-tier LLM completions for planning and decomposition
Cloudflare Agents Adapter bernstein.adapters.cloudflare_agents Spawn agents via npx wrangler dev locally
Codex-on-Cloudflare Adapter bernstein.adapters.codex_cloudflare Run OpenAI Codex inside Cloudflare sandboxes
D1 Analytics bernstein.core.cost.d1_analytics Usage metering, billing tiers, quota enforcement
Vectorize Cache bernstein.core.memory.vectorize_cache Semantic cache for LLM responses using vector similarity
MCP Remote Transport bernstein.mcp.remote_transport Streamable HTTP transport for remote MCP server access
Cloud CLI bernstein.cli.commands.cloud_cmd bernstein cloud subcommands (login, run, status, cost, deploy)

What you need

At minimum:

  • A Cloudflare account (free tier works for Workers AI and basic Workers)
  • A Cloudflare API token with appropriate permissions
  • Your Cloudflare account ID

For the full stack, you also need:

  • An R2 bucket for workspace sync
  • A D1 database for analytics
  • A Vectorize index for semantic caching

See Setup for step-by-step provisioning instructions.