apache/maka

Selection

Select when

• You want a standalone local-first coding-agent workspace rather than a Claude Code extension.
• You need both desktop and terminal interfaces over a shared agent runtime.
• You want model and tool activity retained as a recoverable execution record.
• You need bring-your-own cloud, local, or gateway model connections.
• You want approval-gated filesystem, shell, and network-related agent operations.
• You need session branching, retries, artifacts, search, and crash recovery.
• You want optional MCP, web search, memory, computer use, or catalog skills.
• You need declarative comparisons between Maka and external agent subjects.
• You can accept beta software and changing commands or data formats.
• Your supported environment matches the documented CLI or desktop platform matrix.

Boundaries

Avoid when

• You require a native Claude Code plugin, hook, command, skill, or MCP server specifically for Claude Code.
• You need a production-supported or approved ASF release.
• You require a stable data format and command surface.
• You need the desktop application on Linux or Intel macOS.
• You require a signed and fully supported Windows desktop release.
• You require credentials to be stored in an operating-system keychain.
• You do not want to supply and configure a model provider or local model.
• You cannot grant a Node-based agent workspace access to project files and optional shell execution.
• You require every security-sensitive implementation claim to have been independently audited.
• You need benchmark setup without Docker, pinned Python environments, or machine-local benchmark assets for Harbor or Pier.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so implementation details in SECURITY.md, sandbox code, MCP code, computer-use code, skill policy, tests, and release workflows were not directly inspected.
• Uncertainty: repository metadata and documentation dates extend into 2026 relative to this analysis context, so recency, adoption, issue counts, and release state cannot be independently reconciled.
• The project explicitly states that it is under active development and that commands, data formats, and experimental capabilities may change.
• No Apache release exists yet, and Desktop nightly builds are explicitly for developers and testers rather than production.
• Desktop currently supports Apple Silicon macOS, while Windows is an unsigned preview and Linux and Intel macOS are unsupported.
• The CLI requires Node.js 22.19 or newer, and some graph, evaluation, peer, or media-facing workflows require additional platform dependencies.
• Maka does not bundle a shared model account, so useful operation requires a configured API, local model, supported account, or compatible gateway.
• API keys are stored in a local plaintext credential-vault file protected by the operating-system account boundary rather than an OS keychain.
• The explicit yolo mode grants full file and network access, substantially weakening the default approval boundary.
• No native Claude Code integration is documented; classification as an alternative or companion is an inference from Maka's standalone desktop, CLI, and runtime architecture.
• Optional web search, cloud models, MCP servers, computer use, and external evaluation subjects can expand data exposure and attack surface when configured.
• No independent security audit, privacy assessment, cross-provider quality benchmark, or comprehensive compatibility matrix was supplied.

Filesystem-writing and shell tools must cross a sandbox boundary, and operations outside it require approval by default; runs can be aborted and failures classified. The CLI documents an explicit yolo mode that grants full file and network access. The Electron renderer has no direct Node API and reaches the main process through a typed contextBridge and guarded IPC. Runtime credentials stay outside the runtime package, but the current credential vault is plaintext and relies on OS-account permissions; POSIX installs enforce owner-only modes. Eval documents authenticated result frames, bounded parsing, egress controls, and fail-closed preflight for selected harnesses, but the supplied implementation was not inspected and some documented egress limits remain.