MadAppGang/magus

Selection

Select when

• You want a broad Claude Code plugin suite rather than one isolated command or skill.
• Your team wants project-committed plugin enablement for a shared workflow baseline.
• You need structured implementation, debugging, review, QA, research, or documentation workflows.
• You want semantic code navigation through an MCP-backed search runtime.
• You want independent external-model votes or delegated model sessions from Claude Code.
• You need Claude to operate interactive terminal applications through owned tmux panes.
• You want UI creation and screenshot-based review integrated with development workflows.
• You need to inventory installed skills and manage their persistent context cost.
• You use Bun, TypeScript, Go, or Dingo and want repository-supplied specialist guidance.
• You author nondeterministic coding-agent benchmarks with madbench.

Boundaries

Avoid when

• You need a minimal extension with no hooks, MCP servers, or external executables.
• You cannot allow prompts, code, or images to reach configured external model providers.
• You require every capability to work offline.
• You cannot install required runtimes such as Bun, tmux-mcp, Mnemex, Claudish, or madbench for selected plugins.
• You need a completed independent security or privacy audit before adoption.
• You cannot permit agents to execute commands, inject keystrokes, edit code, or write project settings.
• You require a stable, independently verified Claude Code compatibility matrix.
• You prefer a single-vendor model path without OpenRouter or other external provider routing.
• You need broad Windows support for workflows whose documented prerequisites are tmux-oriented.
• You require benchmark or productivity claims to be independently reproduced.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so manifests, hook implementations, MCP configurations, scripts, tests, CI, and most plugin bodies were not directly inspected.
• Uncertainty: repository timestamps and embedded benchmark dates extend into 2026 relative to this analysis context, so recency, release state, and compatibility claims cannot be independently reconciled.
• Uncertainty: the metadata description says 15 plugins, but the supplied files do not provide a complete authoritative manifest or count for all current plugins.
• Repository-reported productivity, quality, token-cost, and business-value claims were not independently reproduced and may be workload-dependent.
• Several plugins require external binaries, services, credentials, or dependent plugins, so installation effort and failure modes vary by selected capability.
• The development plugin requires Bun and depends on Claudish and Multimodel, while other plugins separately require tools such as Mnemex, tmux-mcp, or madbench.
• Multi-model and vision workflows can transmit prompts, code, findings, or images to configured external providers under their own privacy terms.
• Terminal and provisioning plugins can execute commands, inject keystrokes, interact with databases, install or configure integrations, and write project settings, creating substantial operational risk if misused.
• Claudish is described as auto-approving by default in the top-level README, so delegated sessions may make changes without per-action approval unless configured otherwise.
• No independent security audit, privacy assessment, comprehensive cross-platform matrix, or end-to-end Claude Code compatibility test was supplied.
• The metadata reports 586 open issues and only 10 stars, but issue composition and whether counts are meaningful could not be established from the supplied files.
• Individual plugins may use licenses different from the repository-level MIT license, requiring manifest-level review before redistribution or enterprise adoption.

The suite includes powerful write and execution surfaces: lifecycle hooks can block turns or inject guidance, Mnemex exposes editing and memory tools, setup can write configuration, and tmux-mcp can run commands and inject keystrokes. Claudish inherits the Claude Code environment for credentials and may retrieve credentials from its own configuration, macOS Keychain, or 1Password. Documented safeguards include dry-run provisioning, read-only investigation modes, slot isolation, destructive-command approval guidance, explicit plugin dependencies, and human approval before applying learned directives. However, implementations were not inspected, auto-approve is described as Claudish's default, and no independent audit was supplied.