agentic-stack

agentic-stack is an open-source, portable agent brain layer that lets you maintain a single .agent/ folder — containing memory, skills, and protocols — across multiple coding-agent harnesses without losing context when you switch tools. It supports Claude Code, Cursor, Windsurf, OpenCode, OpenClaw, Hermes, Pi Coding Agent, Codex, Antigravity, and DIY Python loops. The project also ships a local data layer for monitoring agent activity, token/cost estimates, KPI dashboards, and a data flywheel for turning approved runs into training-ready artifacts — all without sending telemetry.

• Portable .agent/ brain — A single folder with four memory layers (working, episodic, semantic, personal), nine seed skills, enforced permissions, and adapters for 10+ harnesses; install once and reuse across projects.
• Harness adapters — Manifest-driven adapter system (adapter.json) for Claude Code, Cursor, Windsurf, OpenCode, OpenClaw, Hermes, Pi, Codex, Standalone Python, and Antigravity; managed via verb subcommands (add, remove, status, doctor, manage).
• Memory review protocol — auto_dream.py stages candidate lessons nightly; host-agent CLI tools (graduate.py, reject.py, reopen.py) require a rationale for every decision, keeping reasoning transparent and auditable.
• Seed skills — Nine built-in skills including skillforge, memory-manager, git-proxy, debug-investigator, deploy-checklist, design-md, data-layer, data-flywheel, and opt-in tldraw canvas; skills use progressive disclosure so only relevant SKILL.md files load.
• Transfer wizard — agentic-stack transfer moves a project brain (preferences, lessons, skills, memory) into another harness via a natural-language TUI, with SHA-256 verification and secret-like content blocking.
• Data layer — Local-only dashboard exports (dashboard.html, daily-report.md) across all harnesses sharing the same .agent/ brain, covering harness events, cron timelines, token/cost estimates, and user-defined categories.
• Data flywheel — Exports approved, redacted runs into trace records, context cards, eval cases, and training-ready JSONL under .agent/flywheel/; model-agnostic and fully local.
• FTS memory search (beta) — Opt-in full-text search over all memory documents using FTS5, with ripgrep and grep fallbacks.
• Cross-platform install — Homebrew formula for macOS/Linux (brew tap codejunkie99/agentic-stack) and a PowerShell installer for Windows; also cloneable directly from GitHub.