gaia

What the foundation provides

GAIA positions itself as the foundation you build your whole application on top of, with the React front end already handled. You start a project with a single command, npx create-gaia@latest my-app, and then run a slash command called /gaia-init to finish the last mile setup. That setup configures the project title and package name, installs dependencies through pnpm and corepack, scaffolds internationalization language files, and wires in Claude skills, plugins, and MCP servers such as React Doctor, Playwright CLI, typescript-lsp, claude-obsidian, and Serena. It also initializes spec-kit so the spec lifecycle hooks are in place before your first feature. The intended audience is engineering teams whose AI generated output is outpacing their review process, engineering leads standardizing how Claude works across an organization, and individuals building apps with Claude. The stated requirements are Node.js 22.22.0 or newer, macOS or Linux with Windows supported through WSL2 because the hooks and CLI need bash and POSIX file paths. The project notes that it is Claude native and treats the model as an engineer to be managed rather than a tool to be hoped with. The stated audience and the explicit requirements together set a clear boundary: GAIA is built for teams that already commit to Claude driven development and can meet the Node and POSIX expectations, not for a casual clone and run scenario.

How conventions get enforced

The core claim of GAIA is that good output comes from enforced conventions, not from a hope that the model infers them. It embeds Karpathy's four coding principles, thinking before coding, simplicity first, surgical changes, and goal driven execution, plus two of its own rules: always use test driven development and always verify your work. Style and correctness are enforced through roughly 1,450 lint rules, strict TypeScript, Prettier, and Knip, which together catch the patterns Claude drifts into, such as complexity creep, architectural shortcuts, and dead code. Bundled skills load on demand when Claude edits a matching file, and a spec authored through /gaia-spec turns into red state Playwright tests before any source is written. Before a merge, a manager agent scans the branch diff for security and quality problems and blocks the merge until no Critical or Important issues remain, with findings tiered as Critical, Important, and Suggestions. A quality gate before commit then requires typecheck, lint, tests, and build to all pass. Guardrails also block destructive moves through a filesystem deny list and pre tool use hooks that reject dangerous commands. The README frames this as treating Claude like an engineer who must be managed, with the rules and gates acting as the review process that keeps unsupervised output from reaching production. By keeping the model inside a managed process instead of trusting it to self regulate, GAIA aims to make AI generated code safe to merge at the pace a team actually works, rather than slowing development to hand review every line.

Token efficiency and debt control

GAIA spends noticeable effort on keeping context small and stopping tech debt from compounding. Rules are scoped so Claude loads only the ones that match what it is editing, and the project defines five tiers of memory, a wiki committed to git, a hot cache for the current session, a cross session handoff, per agent memory, and a global user memory, so the model stops relearning the codebase each session. A periodic knowledge audit sweeps memory, wiki, and autoloaded files for duplication and stale content. Task orchestration runs each phase in a clean subagent context, and Serena gives symbol aware code intelligence instead of grep. For upkeep, /update-deps acts as an autonomous Dependabot that applies codemods and major migrations, /update-gaia pulls new GAIA releases into your project with a three way merge per file, and /gaia-debt works through a tracked backlog of tech debt found during audits. An optional GAIA CI bot runs maintenance against your subscription, capped at five dollars per run, auto merging patch and minor dependency bumps on green and routing bigger changes to review. The net effect is continuous maintenance rather than periodic cleanup. By wrapping the editor in these gates and keeping the model inside a reviewed process, GAIA tries to make AI assisted development something a team can trust as it scales rather than a shortcut that accumulates hidden cost.

Editorial conclusion

GAIA is published under the MIT license, lists Shell as its primary language in the repository, and was last updated on 2026-08-24, with its homepage at gaiareact.com.