Maki Coding Agent

The README describes syntax-tree-based shell permission splitting, opt-in telemetry, SSRF protection, and an explicit --yolo bypass, showing meaningful attention to permissions and external effects. The supplied evidence does not contain the prompt rules, denial paths, or implementation, so least privilege and confirmation do not receive full credit. Providers, environment variables, dynamic providers, persistent memory, and telemetry are disclosed, but transmitted fields, destinations, retention, and deletion guarantees are incomplete. Cargo uses a keyring and largely constrained dependencies; CI checks Nix Git-dependency hash drift, while two Git patches are revision-pinned and justified. No vulnerability scanning, supply-chain policy, or lockfile evidence is supplied. Rewind covers chat only, and the README expressly says code rewind is unavailable. Author, repository, homepage, and MIT copyright attribution agree, justifying full source-attribution credit; the unverified publisher is treated only as unknown.

The README, Cargo workspace, and CI present a broadly consistent Rust/Nix product, with no obvious contradictions in identity, version, or architecture. Dependencies are centrally declared, Git inputs are pinned by tag or revision, Nix CI checks hash drift, and binaries, Cargo, and Nix provide multiple acquisition paths. Credit is limited because the dependency surface is large and includes a release candidate and upstream Git patches. None of the supplied files shows user-facing failure messages, diagnostic guidance, or recovery messages, so failure_messages scores zero; this does not assert that runtime error handling is absent.

The evidence clearly addresses interactive coding, headless stream-JSON operation, ACP editor integration, small terminals, subagents, MCP, skills, plugins, and numerous cloud and local model providers, thoroughly covering audiences and scenarios. Some boundaries are explicit: code execution has time and memory limits, telemetry is opt-in, rewind excludes code, and provider plugins are a work in progress. A systematic language-support matrix, tool authority model, and plugin isolation boundary are not supplied. Commands, shortcuts, configuration switches, and permission-parsing triggers are concrete, but ambiguity, conflict, and false-trigger behavior are not evidenced. Linux, macOS, Windows, Nix, TUI, ACP, and multi-provider instructions justify full environment-fit credit.

The README is clearly divided into benchmarks, features, providers, installation, ACP, documentation, and community, though important details are delegated to external material not included here. Installation comprehensively covers three operating-system families, Nix, Cargo, and prebuilt releases, including destinations and override variables, earning full credit. The maki name, commands, and crate metadata are consistent, but there is no plugin/API stability promise. Configuration, ACP, and command examples are useful, while an embedded FAQ and complete troubleshooting guide are absent. Limitations such as no code rewind, unfinished Lua provider plugins, and predominantly agent-written code are candidly disclosed, but not as a comprehensive limitations inventory. Cargo metadata and the complete LICENSE consistently establish MIT licensing. A version and release artifact are present, but no changelog or compatibility policy is supplied. The named author, email, repository, website, and community offer maintenance paths, but governance, support duration, and succession responsibility are unspecified.

The described TUI, headless stream-JSON output, session resume, search, subagent views, memory, and editor integration give substantial evidence of usable output pathways, but no actual output samples or statically inspectable output contract are included. The index tool, execution pipeline, Lua extensibility, and granular command permissions plausibly add value beyond a basic coding terminal, though most evidence is author-authored description. Token-saving estimates and benchmark pass/cost figures are stated, with the token estimate appropriately labeled as personal usage; however, the linked report and chart contents are not among the supplied files, so methodology, representativeness, and comparison fairness cannot be verified and cost-benefit does not receive a high score.

Several claims point to a benchmark archive, documentation, plugins, and discussions, but those targets are not included. Claims about performance, 60 FPS, memory use, SSRF protection, and permission correctness lack corresponding implementation excerpts, leaving claim traceability thin. Product identity, version, language, author, and license are corroborated across README, Cargo, and LICENSE, while CI corroborates Nix building and hash checking; most feature claims nevertheless remain single-source README assertions. The author separates a personal token measurement with “for my usage” and expressly labels WIP items, the absent code rewind, and the agent-authorship disclaimer. That supports fact/inference separation, but promotional judgments such as “SUPER fast” and performance conclusions without included methodology prevent full credit.