SUCORO
Back to Blog

Sandboxed: How Sucoro Is Extending Gemini CLI for Parallel Agentic Workflows

Behind-the-scenes look at experimental integrations and bleeding-edge agentic infrastructure

Sucoro Team
8 min read
Experimental, Gemini CLI, Agents, Integration

Sandboxed: How Sucoro Is Extending Gemini CLI for Parallel Agentic Workflows


Published by Sucoro — Jun 27 2025


Welcome to Sandboxed, Sucoro's behind-the-scenes look at experimental integrations, deep system customization, and the bleeding edge of agentic infrastructure. This series documents how we push the limits of today's AI tooling, prototype tomorrow's platforms, and share the insights we uncover along the way.


In this inaugural post, we're diving into the Gemini CLI—an open-source command-line interface that wraps Google's Gemini models with agentic behavior, ReAct-style reasoning, and extensible tool support. At Sucoro, Gemini CLI serves as both a testbed and a node in larger multi-agent pipelines. But its default setup is just the beginning.


We've found that customizing Gemini CLI enables a completely new class of workflows—ones that blend local tools, external APIs, and shared memory systems into a composable developer environment. Here's how we're evolving it.


---


🔧 Customizing Gemini CLI the Sucoro Way


🧩 Tool & MCP Server Integration


Gemini CLI supports tools (actions the agent can perform) and Model Context Protocol (MCP) servers—both local and remote. We've leveraged this to integrate everything from text-to-video models (e.g., Veo) to grounding tools like Google Search. These services are configured via JSON and loadable on demand, allowing flexible and modular deployments.


🧠 Contextual Memory with GEMINI.md


Gemini supports scoped memory through GEMINI.md files. We structure ours hierarchically—global, repo-level, and component-specific—allowing the agent to dynamically load the most relevant context based on execution scope. It's a lightweight memory hierarchy with powerful results.


⚠️ Behavioral Modes (YOLO Mode)


We've selectively enabled YOLO mode, granting agents full autonomy without requiring user confirmation. It's powerful but dangerous—so we use it only in sandboxed dev environments, with clearly defined security boundaries.


🛠 Additional Enhancements


  • Visual customization with themes
  • Logging and verbosity toggles
  • Codebase exclusion rules
  • VS Code extensions for live agent management

    • ---


    🔁 Experimental Integration: Claude + Gemini + Shared Graph Memory


    One of our live experiments connects the Claude Code SDK directly to Gemini CLI, using a shared context graph to pass memory and prompt state between them. This allows:


  • Claude to handle fine-grained code reasoning
  • Gemini to orchestrate higher-level task workflows
  • Shared memory state that both can access and update

    • This model fusion is orchestrated via custom MCP services and real-time memory sync logic. Below is a conceptual diagram of how we're wiring it all together.


    mermaid title="Claude Code SDK + Gemini CLI Integration" type="diagram"

    flowchart TD

    subgraph Claude

    CC["Claude Code SDK"]

    CC -->|Context Output| SharedGraph

    end


    subgraph Gemini

    GC["Gemini CLI"]

    SharedGraph -->|ReAct + Tool Use| GC

    end


    SharedGraph["Shared Graph Memory"]

    Tools["External Tools & MCP Servers"]

    GC -->|Tool Calls| Tools

    CC -->|Fine-tune Queries| Tools


    ---


    🧠 Shared Memory Architecture


    Here's how we think about memory across agents:


    mermaid title="Shared Memory Map" type="diagram"

    graph TB

    UserInput(["User Prompt"]) --> InferenceRouter

    InferenceRouter --> Claude

    InferenceRouter --> Gemini


    Claude -->|Write| MemoryGraph

    Gemini -->|Write| MemoryGraph

    MemoryGraph -->|Contextual Recall| Claude

    MemoryGraph -->|Contextual Recall| Gemini


    subgraph "Agent Layer"

    Claude

    Gemini

    end


    MemoryGraph[["Shared Memory Graph"]]

    InferenceRouter{{"Dynamic Agent Orchestration"}}


    ---


    🚀 What's Next in Sandboxed


    The Sandboxed series will continue to explore:


  • Cross-agent orchestration with shared memory
  • Deploying local agent infrastructure
  • Real-world GPT vs Claude behavior comparison
  • Graph-driven LLM coordination

    • Follow along—we're just getting started.


    ---


    Want to contribute or integrate your own SDK into our open experimentation lab? Contact the Sucoro team at [info@sucoro.com](mailto:info@sucoro.com).


    🧪 Built in the Lab. Tested in the Wild.