RandComp Studio, a new local-only desktop application for macOS, has entered early preview with a distinct architectural approach to AI agent orchestration. Rather than relying on a single long-running session that often suffers from context drift, the tool decomposes work into discrete tasks managed by specialized roles. These roles can be backed by different AI command-line toolsβ€”such as Claude Code, Codex, Kimi, or local models via Ollamaβ€”allowing users to mix high-performance models for planning with cheaper or private models for execution.

Local-First Architecture and Privacy Controls

The application operates entirely on the user's machine, with no RandComp cloud service acting as an intermediary. All project files, agent terminals, and coordination logic reside locally, and the app’s server listens exclusively on 127.0.0.1. While the AI tools themselves communicate with their respective providers (Anthropic, OpenAI, etc.) using the user's own accounts, RandComp itself collects only anonymous usage counts. These metrics include app opens, project creations, and MCP connections, but explicitly exclude prompt content, file paths, and project names. Users can disable this telemetry entirely in the settings.

Independent QA and Role Separation

A key feature of RandComp Studio is its insistence on independent verification. The system assigns a separate 'checker' role to verify the output of the 'builder' role. This checker does not see the builder's reasoning chain, forcing it to evaluate the final result against the requirements rather than trusting the builder's self-assessment. This separation of duties is designed to catch blind spots and hallucinations that might slip through when a single agent plans, builds, and reviews its own work. The output is a set of plain files, including plan.md, DESIGN.md, and qa-report.md, which can be edited or backed up manually.

Multi-Interface Access via CLI and MCP

Users can interact with the local agent team through three distinct interfaces: the graphical macOS app, a command-line tool, and the Model Context Protocol (MCP). The randcomp command allows for scripting and remote management via SSH, outputting JSON for automation. Meanwhile, the MCP integration enables other AI assistants, such as Claude Desktop or Code, to query the RandComp team’s status and assign tasks directly from their chat interface. This ensures that the local agent company remains accessible regardless of which tool the developer is currently using.

Key Takeaways

  • RandComp Studio is currently free during early preview but requires a Mac with Apple silicon (M1 or newer) and macOS 11+.
  • The tool supports multiple AI backends, including Claude Code, Codex, Kimi, and local Ollama models, allowing for cost-optimized role assignment.
  • An independent QA agent verifies work without seeing the builder's reasoning, aiming to reduce self-delusion in AI outputs.
  • Data privacy is strictly local; RandComp adds no cloud layer, and telemetry is limited to anonymous event counts that can be disabled.

The Bottom Line

RandComp Studio tackles the 'drift' problem in agentic coding by enforcing a corporate-like separation of duties, which is a necessary evolution for reliable local AI workflows.