Make AI coding agents reliable in any repo — in one command. Persistent memory, a spec-before-code workflow, and quality gates that actually run. Works with Claude Code, Codex, Cursor, and any AGENTS.md tool.
npm create agentik@latest my-app # new project
npm create agentik@latest add # …or add to an existing one
npm create agentik@latest update # refresh framework-owned files later
npm create agentik@latest update -- --layout compact # migrate old installs
npm run agentik:update # after Agentik is installed locally🇩🇪 Deutsche Version: README.de.md
Agents forget context between sessions, drift from conventions, and hallucinate APIs. Prose instructions alone don't fix that — this template combines three mechanisms that do:
- Memory as files —
memory/CONTEXT.md(current state, hard 80-line budget), ADRs inmemory/decisions/,conventions.md,glossary.md. Read at task start, written at defined moments (protocol in AGENTS.md). - Spec-driven workflow — every non-trivial task starts as a spec with checkable acceptance criteria and test cases designed before code, gets executed step-by-step, and ends with an acceptance walk.
- Enforced gates —
pnpm verify(typecheck + lint + test) is the Definition of Done. It's a script, not a suggestion; a Claude Code hook runs it automatically.
Anti-hallucination is rule zero: nothing is imported, called, or referenced
without verifying it exists first (rules/00-verification.md).
# 1. Scaffold (interactive: pick a profile, name it — done)
npm create agentik@latest my-app
cd my-app
# 2. Open with your agent and finish setup
claude # or Codex, Cursor, …
/init-foundation # wires the gates + captures your domain — or say "init"The scaffolder does the structural work for you (template, profile,
.agentik/framework.config.json, git). New projects get a plain package.json
(named after your project, no framework author/repository metadata).
/init-foundation then handles the judgment parts: wiring real gate commands and
capturing your business domain.
Already have a project? Add the framework without touching your code:
cd my-existing-app
npm create agentik@latest add # copies framework files only; never your src/package.jsonPrefer no install step? npx degit Bobblo99/agentik my-app still works
(then run /init-foundation).
The chosen profile (web-frontend, fullstack, generic) keeps the
matching rules/skills and parks the rest in .agentik/disabled/ (nothing
deleted). Switch profile or toggle modules anytime with /configure — the
active set lives in .agentik/framework.config.json. Then:
/write-spec add user login # agent plans, you approve
/execute-spec specs/2026-...md # agent builds, checks off, verifies
/sync-memory # end-of-session memory update
Codex and other tools follow the identical workflow via AGENTS.md — the slash commands are convenience wrappers, not requirements.
After adopting the framework, update it from any project root:
npm create agentik@latest update -- --dry-run
npm create agentik@latest update
npm create agentik@latest update -- --layout compact # migrate classic installsupdate refreshes framework rules, core skills, commands, profiles, scripts,
and docs while preserving memory, specs, custom rules, custom skills,
configuration, package.json, and application code. Disabled modules stay
parked, and the project's package-manager adaptation stays intact.
Installed projects get create-agentik as a devDependency plus simple scripts:
npm run agentik:update
npm run agentik:checkNew CLI installs use the compact layout so the project root stays readable.
AGENTS.md remains in the root because agents discover it there; it is only a
small bridge to the real framework files. Agentik's own README, changelog,
contributing guide, and license live under .agentik/ in compact installs;
the root belongs to the user's project.
AGENTS.md root bridge for AGENTS.md-compatible tools
CLAUDE.md root bridge for Claude Code
.agentik/AGENTS.md single source of truth
.agentik/framework.config.json
.agentik/skills/ agent-neutral skills for Codex, Cursor, Claude, others
.agentik/claude/ Claude compatibility commands, settings, and skill mirror
.agentik/cursor/ Cursor rule mirrors
.agentik/rules/ binding standards (+ rules/custom/ for your domain)
.agentik/disabled/ parked modules — never read, brought back by /configure
.agentik/specs/ one file per task: plan, test cases, acceptance, archive/
.agentik/memory/ CONTEXT.md · domain.md · decisions/ · conventions.md · glossary.md
.agentik/profiles/ web-frontend / fullstack / generic
.agentik/scripts/ verify.sh · check-framework.sh
.agentik/docs/ orchestration, MCP guidance, adoption guide
.agentik/README.md Agentik reference README, not the user's project README
.agentik/CHANGELOG.md Agentik changelog
.agentik/LICENSE Agentik license
Older classic installs remain supported. Move one to compact layout with:
npm create agentik@latest update -- --layout compact- One source of truth. AGENTS.md ≤ ~150 lines, imperative, commands first; detail lives in skills/rules loaded on demand. Long context files measurably hurt agent performance.
- Checkable over vague. "
pnpm testpasses" beats "code is clean" — everywhere: rules, specs, memory. - Enforce, don't hope. Rules that matter exist twice: as text and as an executable check (verify script, hooks).
- Modular stack. The core is stack-agnostic; React/Next.js lives in a removable profile.
The memory system isn't a tidy folder — it's a human-readable implementation of
established findings: small/budgeted working memory because long context
degrades (Lost in the Middle),
externalized files as tiered memory (MemGPT),
and a working/episodic/semantic/procedural split that mirrors cognitive-agent
architecture (CoALA) with an
encode→retrieve→consolidate cadence (Generative Agents).
Full rationale and the file→memory-type mapping: docs/memory-model.md.
See docs/adopting.md for existing-project adoption and maintenance.
- Source and template repository: Bobblo99/agentik
- npm scaffolder:
create-agentik - The repository changelog describes Agentik itself. Adopted projects should start their own changelog for application changes.
MIT — see LICENSE.