Roo Code Review: Shutdown, Security, and Alternatives

This Roo Code review starts with the fact many search results bury: the original Roo Code extension shut down on May 15, 2026. Its GitHub repository is archived, version 3.54.0 was the final release, and the former company no longer maintains the VS Code agent.

That does not erase why Roo Code became popular. It offered open-source code, flexible model providers, custom agent modes, local-model support, MCP tools, and approval controls. But a frozen coding agent that can read files and run commands is not a sensible new installation. Existing users should migrate to an actively maintained successor.

Roo Code Review: Quick Verdict

Roo Code was one of the most configurable open-source coding agents for VS Code. Its Code, Architect, Ask, and Debug modes separated implementation from planning and diagnosis, while custom modes let teams narrow prompts and tool permissions for specialist work. Users could bring API keys, choose local models, and approve file or terminal actions.

The current verdict is straightforward: do not adopt the original extension for new work. The official repository is read-only, the hosted services and paid plans are no longer the buying decision, and future model APIs, VS Code changes, or security discoveries will not receive fixes from the original project.

Zoo Code is the closest continuity option because it is a community fork of Roo Code. Cline is the safer default for users who value a larger, actively maintained project and can accept a different workflow. Cursor, OpenHands, and other agents fit buyers who want a full IDE or a broader autonomous platform.

What Was Roo Code?

Roo Code was a free, Apache-2.0-licensed VS Code extension derived from Cline. It translated natural-language requests into actions such as reading and writing project files, executing shell commands, browsing the web when supported, and calling external tools through the Model Context Protocol.

The extension was an orchestration layer rather than a model. Users connected providers such as Anthropic, OpenAI, OpenRouter, or local inference through Ollama and LM Studio. Relevant prompts and project context could be sent to the selected provider, so privacy depended on both Roo Code's settings and that provider's terms.

Its main differentiator was modes. Code handled general implementation, Architect focused on planning, Ask answered questions, and Debug supported diagnosis. Custom modes, project rules, auto-approval settings, and semantic codebase indexing made the extension unusually adaptable for developers willing to configure it.

Why Did Roo Code Shut Down?

The official GitHub notice says the Roo Code extension shut down on May 15, 2026 and directs users to Zoo Code and Cline. The repository was archived on the same date, making it read-only. It also tells former paying users to contact Roo Code about billing.

The shutdown changes the product evaluation more than any historical feature comparison. An archived extension may continue to launch, but it no longer has a maintainer responsible for compatibility, dependency updates, provider changes, or newly disclosed vulnerabilities. An agent with access to source code and a shell needs an active patch path.

The old code remains available under Apache 2.0, which allowed community forks to continue development. Open-source availability preserves inspectability and reuse; it does not make the abandoned binary operationally current.

Roo Code Features and Historical Strengths

Multi-Model and Local Inference

Roo Code let users choose models rather than forcing one bundled provider. That made it possible to route simple work to a cheaper model, reserve a stronger model for difficult changes, or run compatible local models when code could not leave the machine.

This flexibility also moved work onto the user. Teams had to manage keys, rate limits, model compatibility, context windows, and provider-specific data terms. A free extension did not mean free inference.

Custom Modes and Project Rules

Modes gave different tasks distinct instructions and tool access. A planning mode could stay read-only, while a coding mode could edit files after approval. Custom modes and .roorules files helped teams encode repository conventions and repeatable workflows.

That separation was useful but not a security boundary by itself. The model, extension tools, VS Code environment, repository content, and user approvals all remained part of the same risk surface.

Terminal, Browser, MCP, and Indexing

Roo Code could execute terminal commands, access the web through supported models, and connect to MCP servers. Its semantic index used embeddings and a Qdrant vector database to find related code by meaning rather than exact keywords.

Those capabilities made the agent effective across larger repositories. They also increased setup and governance work. External MCP servers can introduce code, credentials, or network access; indexing can send code-derived content to an embedding provider; and terminal execution can turn a bad suggestion into a real system change.

Roo Code Pricing and Data Handling

The original extension was free and open source, but users normally paid their chosen model provider for tokens. Codebase indexing could add embedding and vector-storage costs. Historical Roo Cloud pricing is now irrelevant because the original product and services have shut down.

Roo Code's final privacy policy said prompts and relevant files were sent to the chosen AI provider, API keys stayed local except when used with that provider, and anonymous telemetry was enabled by default with an opt-out. It also said Roo Code did not store prompts or train models on user data.

Those statements did not override provider policies. Local inference reduced third-party model exposure, but web tools, MCP servers, telemetry, package managers, and source-control integrations could still transmit information.

Roo Code Security: What the Record Shows

Roo Code's archived repository lists multiple 2025 security advisories, including issues involving auto-approved commands, protected workspace files, MCP configuration, and unintended file disclosure. The maintainers published patched versions for the listed extension vulnerabilities, and the final 3.54.0 release came after those fixes.

The record should not be read as proof that Roo Code was uniquely unsafe. Agentic coding tools deliberately combine untrusted text, file access, command execution, and external services, creating a broad attack surface. Public advisories also show that maintainers found and fixed concrete problems.

The present concern is lifecycle. Because the original repository is archived, new flaws will not be patched there. Existing users should export rules and configuration, remove stored credentials, uninstall the abandoned extension after migration, and review any auto-approved commands or writable paths. SD's AI agent security guide provides a broader containment checklist.

Best Roo Code Alternatives

Zoo Code for the Closest Migration

Zoo Code describes itself as a community fork created to continue Roo Code development. Its repository offers a Roo-to-Zoo migration guide and says former Roo contributors are involved. This makes it the most direct option for users who depend on Roo-style modes, provider settings, and project rules.

The tradeoff is governance. A community fork can preserve features and ship fixes, but teams should review maintainer activity, signed releases, security reporting, extension permissions, and migration behavior before trusting it with sensitive repositories.

Cline for Active Maintenance

The archived Roo repository explicitly names Cline, the project from which Roo originated. Cline is actively maintained across IDE, CLI, and SDK surfaces and supports bring-your-own-key or usage-based inference. Its workflow is not identical, so test rules, approvals, MCP configuration, and cost on representative tasks. SD's Cline review covers current pricing, privacy, and security in detail.

Cursor or OpenHands for a Different Workflow

Cursor is a full AI-native editor rather than an extension added to stock VS Code. It is a better fit for users who prefer an integrated commercial experience over maximum provider flexibility; see SD's Cursor AI review.

OpenHands is relevant when the goal is a broader software-development agent that can run in isolated environments rather than a Roo-like editor panel. SD's OpenHands review and top AI coding agents guide compare those workflow categories.

How to Migrate from Roo Code Safely

Inventory custom modes, .roorules, provider settings, MCP servers, ignore files, approval policies, and any cloud-backed state. Remove secrets before copying configuration, then map each capability to the replacement rather than importing everything blindly.

Install the successor in a disposable repository first. Test read boundaries, writes, command approval, rollback, MCP behavior, local-model routing, and cost with a fixed task set. Confirm that ignored and sensitive files stay outside the agent's context.

After validation, revoke obsolete keys, uninstall Roo Code, remove unused local state, and document the new maintenance and security-update process. Keep human approval for destructive commands, credential access, publishing, deployment, and production changes.

Conclusion

This Roo Code review finds an influential open-source coding agent whose original product is no longer a responsible new choice. Its multi-model architecture, custom modes, local inference, project rules, MCP support, and approval controls explain its lasting search interest, but the May 2026 shutdown outweighs those historical strengths.

Existing users should migrate while their configuration is still understandable and testable. Choose Zoo Code for the closest continuity, Cline for active open-source maintenance, or a different agent category when the workflow itself should change. Whatever the replacement, narrow permissions, verify provider data terms, and require human review for consequential actions.