Are Gumloop Agents Built on Claude Code or a Provider Agent SDK?

Last updated: May 19, 2026

No. Gumloop Agents run on our own in-house agent orchestration loop - not Claude Code, the Claude Agent SDK, the OpenAI Agents SDK, LangChain, or LangGraph. The loop is model-agnostic and routes to Anthropic, OpenAI, or Google depending on the model selected on the agent.

The Short Answer

Gumloop Agents are powered by a custom orchestration system we built from scratch. The agentic loop — LLM calls, tool execution, streaming, context management, and message formatting — is entirely Gumloop's own stack.

We do not wrap any provider-specific agent SDK. That means switching the model an agent uses doesn't change the runtime behavior, and you're never locked into one provider's agent abstractions.

Which Models the Loop Routes To

The same loop calls every supported provider through a unified internal interface. Today that includes:

  • Anthropic — Claude 4.5 Opus / Sonnet / Haiku, Claude 4 Opus / Sonnet, Claude 3.7 Sonnet

  • OpenAI — GPT-5.2 / 5.1 / 5 / 5 Mini / 5 Nano, GPT-4.1 / 4.1 Mini / 4.1 Nano, o3, o4-mini

  • Google — Gemini 3 Pro, Gemini 3 Flash, Gemini 2.5 Pro, Gemini 2.5 Flash

  • xAI — Grok 4, Grok 3, Grok 3 Mini

  • Others — DeepSeek, Mistral, Perplexity Sonar, Meta LLaMA, plus Azure OpenAI for customers with their own Azure deployment

You pick the model in your agent's Model Config; the orchestration layer underneath is identical regardless of choice.

How Tool Calling Works

Tool calling on Gumloop Agents uses two mechanisms, both driven by our own orchestrator:

  • Each provider's native function-calling API for tool invocation — not a third-party agent SDK abstraction.

  • Model Context Protocol (MCP) for external integrations, including Gumloop's 50+ pre-built MCP servers and any custom MCP server you connect.

For OpenAI and Anthropic models, custom MCP servers can run natively (the provider connects to your MCP server directly). For Google and a few others, Gumloop's backend acts as the MCP connector and executes tool calls on the model's behalf. Either way, the outer loop — deciding when to call a tool, what to do with the result, how to stream the response — is Gumloop's code, not a vendor SDK.

Why This Matters

  • Model-agnostic by design. Switching providers on an agent doesn't require rebuilding anything. The agent definition, tools, skills, and conversation history all stay the same.

  • No provider lock-in. Gumloop is not dependent on any single vendor's agent framework continuing to exist or remaining backward-compatible.

  • Fine-grained control. Credit tracking, context management, error handling, parallel tool calls, and streaming are all owned by Gumloop and tuned for the platform's needs.

Still Need Help?

If you have a deeper architecture or security question, reach out to support at support@gumloop.com.

Related Docs